International Series in Operations Research & Management Science Sharan Srinivas Suchithra Rajendran Hans Ziegler Editors Supply Chain Management in Manufacturing and Service Systems Advanced Analytics for Smarter Decisions International Series in Operations Research & Management Science Founding Editor Frederick S. Hillier Stanford University, Stanford, CA, USA Volume 304 Series Editor Camille C. Price Department of Computer Science, Stephen F. Austin State University, Nacogdoches, TX, USA Associate Editor Joe Zhu Foisie Business School, Worcester Polytechnic Institute, Worcester, MA, USA More information about this series at http://www.springer.com/series/6161 Sharan Srinivas • Suchithra Rajendran • Hans Ziegler Editors Supply Chain Management in Manufacturing and Service Systems Advanced Analytics for Smarter Decisions Editors Sharan Srinivas Department of Industrial and Manufacturing Systems Engineering, College of Engineering, Department of Marketing, Trulaske College of Business University of Missouri Columbia, Missouri, USA Suchithra Rajendran Department of Marketing, Trulaske College of Business, Department of Industrial and Manufacturing Systems Engineering, College of Engineering University of Missouri Columbia, Missouri, USA Hans Ziegler School of Business, Economics and Information Systems University of Passau Passau, Germany ISSN 0884-8289 ISSN 2214-7934 (electronic) International Series in Operations Research & Management Science ISBN 978-3-030-69264-3 ISBN 978-3-030-69265-0 (eBook) https://doi.org/10.1007/978-3-030-69265-0 © Springer Nature Switzerland AG 2021 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG. The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland To the memory of my father, B. Srinivasan, who always believed in me and supported all my choices in life. —Sharan Srinivas To my husband, parents and my late grandmother (Dr. A.B Vasanthalakshmi) —Suchithra Rajendran To Sarah Madeleine and Ulrike —Hans Ziegler Preface Management of supply chains has evolved rapidly over the last few years owing to the inception of Industry 4.0, where digitization using Internet of things (IoT), advanced robotics, and sensors have taken precedence. The traditional supply chain systems are no longer efficient as these transformations have led to dynamic and interconnected systems that require robust management capabilities. Moreover, today’s customers are being exposed to numerous products and services through the Internet, leading to high expectations requiring superior supply chain systems with cost-effective, responsive, and sustainable capabilities. Nevertheless, digital transformation, both in manufacturing and service industries, has led to large volumes of real-time data which can be leveraged to obtain actionable insights and drive profitable supply chain decisions. This book focuses on providing an overview of current trends in supply chain as well as publishing state-of-the-art original research work dealing with advanced analytical models (predictive and prescriptive analytic models) for the design, planning, and operation of supply chains in the era of digitization and Industry 4.0. It intends to empower supply chains with smarter decisions at all levels and stages. The key characteristics of this book are as follows: 1. Covers recent trends, developments, and applications in supply chain management 2. Includes a wide collection of analytical methodologies for optimizing key supply chain decisions 3. Bridges the theory–practice gap in supply chain management 4. Designed, organized, and edited considering non-experts 5. Holistic with contributions from leading academicians and industry practitioners 6. Unified single-source guide 7. Versatile reference book for students, researchers, educators, and practitioners, alike. vii viii Preface The vital areas of supply chain management empowered with smart decisions include: 1. 2. 3. 4. 5. 6. 7. 8. Supply chain network design Logistics and distribution Process optimization Product life cycle management Visibility Risk management Sustainability Customer feedback Book Overview This book presents chapters from well-recognized international authors across academia and industry to facilitate holistic and up-to-date knowledge in supply chain management. The chapter entitled “An Overview of Decisions, Performance and Analytics in Supply Chain Management” introduces the readers to the concept of supply chain management and its various facets. Further, it aims to familiarize the readers with the frequently used terminologies to help them navigate the remaining chapters. The chapter entitled “Intelligent Digital Supply Chains” covers the current trends in digital supply chain management such as intelligent visibility, digitization, and blockchain technology. It sets the context for the research works presented in the following chapters. Further, to address concerns in both product and service supply chains, the book provides novel research work in both the sectors. While chapters entitled “Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design,” “Supply Chain Risk Management in Indian Manufacturing Industries: An Empirical Study and a Fuzzy Approach,” and “Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces” are related to the manufacturing industry, the remaining chapters focus on the service sector. Supply chain management involves every decision taken to optimize the flow of products, funds, and information in an organization. These decisions can be better perfected by understanding them at one of the three decision levels – strategic, tactical, and operational. Strategic decisions are long-term decisions usually taken by the senior management and mainly include decisions on supply chain network design, production and sourcing, and information technology. These decisions need to be carefully accounted for their long-term impact and future uncertainties. The chapters entitled “Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design,” “Supply Chain Risk Management in Indian Manufacturing Industries: An Empirical Study and a Fuzzy Approach,” “Improving Service Supply Chain of Internet Services by Analyzing Online Customer Reviews,” and “Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces” address some of the critical strategic issues in a supply chain. On the other hand, Preface ix tactical decisions are made every month or quarter and are related to supply chain planning activities like purchasing, production and planning, transportation, inventory management, and distribution. The chapters entitled “An Integrated Problem of Production Scheduling and Transportation in a Two-Stage Supply Chain with Carbon Emission Consideration” and “A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for Improving Pharmaceutical Service” prescribe the best course of action for tactical decision-making. Finally, operational decisions are related to determining the day-to-day tasks needed to satisfy individual customer orders. They involve activities like allocating inventory, assigning logistics, updating delivery dates, and placing replenishment orders in response to customer orders. The chapters entitled “Pro-active Strategies in Online Routing” and “Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces” provide insights to make effective real-time operational decisions. While this book is organized based on the decision level each chapter addresses, the chapters can also be understood by the methods they employ. The chapters entitled “Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design,” “An Integrated Problem of Production Scheduling and Transportation in a two-Stage Supply Chain with Carbon Emission Consideration,” Pro-active Strategies in Online Routing,” and “Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces” leverage mixed-integer linear programming (MILP) techniques to obtain the optimal values for the decision variables in a supply chain, given practical constraints. Further, the chapters entitled “Proactive Strategies in Online Routing” and “Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces” integrate heuristics into their MILP formulations to quality solutions more quickly. On the other hand, the chapter entitled “A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for Improving Pharmaceutical Service” evaluates the effectiveness of integrating new strategies to improve last-mile delivery in supply chain using discrete event simulation modeling. The chapter entitled “Supply Chain Risk Management in Indian Manufacturing Industries: An Empirical Study and a Fuzzy Approach” employs fuzzy cognitive map (FCM) to foresee and manage SC risks. Finally, the chapter entitled “Improving Service Supply Chain of Internet Services by Analyzing Online Customer Reviews” uses text analytics techniques like bigram and trigram analysis on online customer reviews to gain insights to make better strategic decisions. Chapter Summaries The chapter entitled “An Overview of Decisions, Performance and Analytics in Supply Chain Management” introduces the readers to the concept of supply chain management and further provides numerous case studies to help them understand its various facets. Particularly, this chapter focuses on decision levels, enablers, x Preface drivers, and varied goals associated with supply chains. Moreover, this chapter also familiarizes the readers with analytics, setting the stage for the book. The chapter entitled “Intelligent Digital Supply Chains” reports on how modern supply chains are leveraging Industry 4.0 technologies to gain end-to-end visibility and intelligence in their supply chains. Further, it discusses planning, execution, evaluation, monitoring, risk management, and opportunities from the perspective of digitized supply chains. This chapter is vital for both academic professionals who want to understand the tools employed by industry practitioners and industry personnel who need to catch up with the latest trends. The chapter entitled “Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design” proposes a product life cycle optimization model for a closed-loop supply chain network design. In environmentally concerning times, a closed-loop supply chain capable of reclaiming and reusing post-consumer materials to reduce wastes and dependence on raw materials is crucial. Further, remanufacturing is efficient and profitable. This chapter proposes an integrated multi-period optimization model to design the closed-loop supply chain for OEMs. The authors explicitly model the optimal collection of remanufactured products through suppliers in a dynamic manner across multiple time periods over the product life, accounting for demand, quality, and remanufacturability of returned products. The model is applied to a realistic case study of Apple’s iPhone 7 for a product life cycle of 8 years. The chapter entitled “Supply Chain Risk Management in Indian Manufacturing Industries: An Empirical Study and a Fuzzy Approach” seeks to predict supply chain risks via early warning signals and implement appropriate mitigation strategies. With many organizations solely focusing on efficiency to win today’s fierce competition, supply chains have become more vulnerable to risks. The authors strive to address this issue using a twofold method. In the first part, the critical factors of supply chain risk management are understood by developing a framework, and the views of the practicing managers about risks perceived in their organization are captured using an empirical study. Second, this information is used to develop a fuzzy cognitive map to identify all plausible risks in the future, given a risk observed from a point in time, and suggest proactive mitigation strategies for practicing managers. The chapter entitled “Improving Service Supply Chain of Internet Services by Analyzing Online Customer Reviews” proposes a methodology to leverage online customer reviews using text analytics to improve service quality and customer satisfaction. It specifically focuses on Internet service providers (ISPs), as the service sector is gaining more attention, and the wireless communications industry has become highly competitive. The chapter seeks to extract the current strengths, weaknesses, opportunities, and threats (SWOT), along with their corresponding root causes for several leading ISPs by exploring consumer reviews using text analytics. The proposed approach consists of four different stages, bigram and trigram analyses, topic identification, SWOT analysis, and root cause analysis (RCA), and provides compelling managerial insights. This chapter is critical, as all businesses can benefit from learning to leverage their online customer reviews. Preface xi The chapter entitled “An Integrated Problem of Production Scheduling and Transportation in a Two-Stage Supply Chain with Carbon Emission Consideration” provides an integrated logistics and scheduling problem with the consideration for carbon emissions in a supply chain system. In today’s highly competitive market, companies strive to capitalize on every opportunity across supply chain stages. Further, while doing so, sometimes they ignore the resultant carbon emissions, leading to growing concerns that mandate focus on carbon emissions as well. This chapter addresses this challenge by providing a solution that integrates three issues: selection of subcontractors, scheduling of jobs, and scheduling of logistics with carbon emission consideration using mixed-integer linear programming (MILP). The chapter entitled “A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for Improving Pharmaceutical Service” reports on the effects of employing drone integrated delivery service at the pharmacy. Drones have the potential to decrease labor costs, maintenance costs, and delivery times, and pharmacy offers a unique opportunity for possible drone delivery as quick delivery can better serve the community. This chapter uses discrete-event simulation to compare drone-only, truck-tandem, and truck-only delivery methods on a variety of scenarios to provide the information needed for municipalities to determine the validity of using drones for pharmaceutical deliveries. The chapter entitled “Pro-active Strategies in Online Routing” summarizes the ideas behind recent approaches to efficiently control urgent delivery processes in real time. With growing customer expectations, businesses periodically face urgent requests that need to be serviced the same day. Further, since these requests are usually unknown and will occur dynamically, the transportation plan that is already in execution must be continuously adapted in real time by applying suitable optimization approaches. This chapter reports on these approaches that control urgent deliveries by minimizing the total weighted request response times in order to minimize resulting customer inconveniences. This chapter is specifically vital for logistics service providers as more businesses seek to improve their supply chain’s responsiveness to accommodate express deliveries. The chapter entitled “Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces” presents prescriptive analytics for dynamic real-time scheduling (DRTS) of the diffusion furnace. In the semiconductor industry, the capability to meet delivery commitment and have shorter cycle time is the most critical challenge in facing global competition. Particularly, wafer fabrication processes involve complex operations, and effectively scheduling them is crucial. This chapter addresses scheduling in diffusion operation, the lengthiest wafer fabrication process involved, by proposing a mathematical model. The chapter proposes seven different apparent tardiness cost (ATC)-based greedy heuristic algorithms (GHA) for the same. Columbia, Missouri, USA Passau, Germany Sharan Srinivas Suchithra Rajendran Hans Ziegler Acknowledgments First and foremost, we would like to thank the authors as this book would not have been possible without their timely and novel contributions. We also acknowledge their efforts in preparing concise chapters for addressing recent and practically relevant supply chain issues using advanced analytical methodologies. Besides, we would also like to thank the authors for promptly revising the chapters based on the reviewer’s comments. We want to express our sincere appreciation to the editorial assistant, Surya Ramachandiran, for his outstanding assistance in several preparatory aspects of this edited book, especially for typing the first chapter, reviewing the language style, and suggesting insightful edits to improve the overall quality of the chapters. We want to thank Christian Rauscher, senior editor of Business, Operations Research & Information Systems at Springer, for his guidance from the conception to the completion of this edited book. We also acknowledge the timely help and information provided by Sayani Dey, production editor at Springer, during different book preparation stages. Finally, we would like to thank our families for their encouragement and support in our research endeavors. xiii Contents An Overview of Decisions, Performance and Analytics in Supply Chain Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . Sharan Srinivas, Suchithra Rajendran, and Hans Ziegler Intelligent Digital Supply Chains . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . Raghav Jandhyala Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . Aswin Dhamodharan and A. Ravi Ravindran 1 19 65 Supply Chain Risk Management in Indian Manufacturing Industries: An Empirical Study and a Fuzzy Approach .. . . . . . . . . . . . . . . . . . . . 107 V. Viswanath Shenoi, T. N. Srikantha Dath, and Chandrasekharan Rajendran Improving Service Supply Chain of Internet Services by Analyzing Online Customer Reviews . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 147 Suchithra Rajendran and John Fennewald An Integrated Problem of Production Scheduling and Transportation in a Two-Stage Supply Chain with Carbon Emission Consideration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 165 Bobin Cherian Jos, Chandrasekharan Rajendran, and Hans Ziegler A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for Improving Pharmaceutical Service . . . . . . . .. . . . . . . . . . . . . . . . . . . . 185 Alexander Jackson and Sharan Srinivas Pro-Active Strategies in Online Routing.. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 205 Stefan Bock Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 241 M. Vimala Rani and M. Mathirajan xv Contributors Stefan Bock Business Computing and Operations Research, Schumpeter School of Business and Economics, University of Wuppertal, Wuppertal, Germany Aswin Dhamodharan Tesla Motors, San Carlos, CA, USA John Fennewald Department of Industrial and Manufacturing Systems Engineering, University of Missouri, Columbia, MO, USA Department of Marketing, University of Missouri, Columbia, MO, USA Alexander Jackson Department of Industrial and Manufacturing Systems Engineering, University of Missouri, Columbia, MO, USA Raghav Jandhyala SAP, Tempe, AZ, USA Bobin Cherian Jos Department of Mechanical Engineering, Mar Athanasius College of Engineering, Kothamangalam, Kerala, India M. Mathirajan Department of Management Studies, Indian Institute of Science, Bangalore, India Chandrasekharan Rajendran Department of Management Studies, Indian Institute of Technology Madras, Chennai, India Suchithra Rajendran Department of Industrial and Manufacturing Systems Engineering, University of Missouri, Columbia, MO, USA Department of Marketing, University of Missouri, Columbia, MO, USA A. Ravi Ravindran Pennsylvania State University, State College, PA, USA T. N. Srikantha Dath Department of Mechanical and Manufacturing Engineering, M S Ramaiah University of Applied Sciences, Bengaluru, Karnataka, India Sharan Srinivas Department of Industrial and Manufacturing Systems Engineering, University of Missouri, Columbia, MO, USA Department of Marketing, Trulaske College of Business, University of Missouri, Columbia, MO, USA xvii xviii Contributors M. Vimala Rani Vinod Gupta School of Management, Indian Institute of Technology, Kharagpur, Kharagpur, India V. Viswanath Shenoi Department of Computer Science and Engineering, Amrita College of Engineering and Technology, Nagercoil, Tamil Nadu, India Hans Ziegler School of Business, Economics and Information Systems, University of Passau, Passau, Germany An Overview of Decisions, Performance and Analytics in Supply Chain Management Sharan Srinivas, Suchithra Rajendran, and Hans Ziegler Businesses today face major challenges in terms of greater competition and increased customer expectations from the global market. With advances in logistics and information technology, today’s customers are exposed to abundant products and services offered worldwide. Thus, businesses aim for competitive advantage and product/service differentiation to stay relevant and profitable. They strive to build robust supply chains that can help them deliver the right product/service more quickly and economically than their competitors. The focus of this book is to provide an overview on the current trends in supply chains as well as present advanced analytical models to optimize the design, planning and operation of supply chains. This chapter discusses the concept of supply chain management, various levels of supply chain decisions and their impacts, drivers and enablers of a supply chain, types of supply chain, and introduces the role of analytics in supply chain. Further, this chapter also presents relevant case studies to help readers better understand various aspects of supply chain management and their importance. Finally, this chapter links the various supply chain problems addressed in this book to the key decision levels and analytical methods, thereby setting the stage for the readers. S. Srinivas () · S. Rajendran University of Missouri, Columbia, MO, USA e-mail: SrinivasSh@missouri.edu H. Ziegler University of Passau, Passau, Germany © Springer Nature Switzerland AG 2021 S. Srinivas et al. (eds.), Supply Chain Management in Manufacturing and Service Systems, International Series in Operations Research & Management Science 304, https://doi.org/10.1007/978-3-030-69265-0_1 1 2 S. Srinivas et al. 1 Overview of Supply Chain A supply chain contains all stakeholders and activities involved in completing a customer’s order, be it a product or a service. Supply chains are not just restricted to the suppliers and manufacturers, but any stage directly or indirectly involved, including transporters, distributors, retailers, and even end-customers. Further, these stages may even be located in different countries across the world for a company with a global supply chain footprint. For example, the journey from coffee bean to a beverage at Starbucks involves such a global supply chain, where coffee beans and related items are brought from around the world and consigned to Starbucks’s 16,700 retail stores to serve over 50 million buyers across 51 different countries each week. A single cup of coffee at Starbucks, from the coffee bean, milk, sugar to the paper cup, can be dependent on as many as 19 countries, connecting some of the poorest countries in the world to the richest. A formal definition of supply chain is given by Ravindran and Warsing (2016, p. 2) using two components: (i) “a series of stages (e.g., suppliers, manufacturers, distributors, retailers, and customers) that are physically distinct and geographically separated at which inventory is either stored or converted in form and/or in value.” (ii) “a coordinated set of activities concerned with the procurement of raw materials, production of intermediate and finished products, and the distribution.” Supply chains tend to be highly dynamic; in addition to product movement, they also involve the flow of information and funds between different stages. For example, e-commerce websites, such as Amazon.com, adopt a series of interrelated activities, as shown in Fig. 1, to satisfy customer orders. The products from the thirdparty sellers are bought and shipped to one of the company’s fulfillment centers present worldwide, via air hubs, port facilities, and cross-docks. Cross-docks are places where goods from inbound transport are removed and then directly loaded onto an outbound carrier, to facilitate logistics efficiency. Usually, the fulfillment centers hold required inventory levels predicted by analytical algorithms to enable express deliveries. The fulfillment centers not only act as a warehouse but also host facilities to package products and prepare them for delivery when needed. Following customer orders, the products are usually moved from the fulfillment center to the nearest sortation center, where they are segregated based on ZIP codes. Consequently, they are transported to their appropriate delivery stations, where the products are prepared for their last-mile delivery. As demonstrated in Fig. 1, real-world supply chains are usually not linear but complex convergent and divergent networks as a manufacturer may source from multiple vendors and then supply to numerous distributors. Further, the flow can happen in both directions and may be controlled by one or more intermediate stages. Finally, through the whole process, along with the product, both information and funds constantly flow to make the supply chain efficient. Thus, most supply chains tend to be complex networks or webs needing holistic management strategies An Overview of Decisions, Performance and Analytics in Supply Chain Management 3 Fig. 1 Stages of E-Commerce Supply Chain to function effectively. The Association for Operations Management (APICS) defines supply chain management as “the design, planning, execution, control, and monitoring of supply chain activities with the objective of creating net value, building a competitive infrastructure, leveraging worldwide logistics, synchronizing supply with demand, and measuring performance globally” (Blackstone 2010, p. 148). 2 Supply Chain Decision Levels While managing a supply chain, numerous decisions need to be taken regarding the flow of materials, information, and funds. The various decisions taken in a supply chain fall under one of the three levels, namely, strategic, tactical, and operational, based on how frequently that decision is taken and duration over which its impact is experienced. 4 S. Srinivas et al. 2.1 Strategic Strategic decisions primarily determine how a supply chain is designed and who will be the partners for the upcoming years. Unlike other decisions, strategic decisions have a significant long-term impact, and making sudden changes is generally both expensive and not feasible. However, both the customer demands and market are ever-changing, making it crucial for the companies to carefully account for the uncertainties tied with the future before making any strategic decisions. These decisions majorly decide the structure of the supply chain, method for material procurement, strategy for allocating resources, and processes undertaken at each stage. Broadly all the strategic decisions taken in a supply chain fall under one of the three categories—network design, production and sourcing, and information technology. First, in network design, decisions regarding the number and type of facilities needed, their geographic locations, and their production and storage capacities are made. Other strategic decisions regarding the mode of transport between these facilities also fall into this category. Second, in production and sourcing, decisions on making or buying (whether to outsource or conduct the activity in-house) at each stage of a supply chain are taken. Moreover, decisions on the selection of vendors, sub-contractors, and other alliances also belong to this category. Finally, strategic decisions can also be related to managing information technology infrastructure. Decisions like what type of information systems are needed, and whether to develop them internally, buy the related commercially available version or employ the freely available open-source alternatives are taken here. An American multinational food and beverage corporation, PepsiCo, is an apt example of how effective strategic decisions can better satisfy customers. While PepsiCo’s best-known products were carbonated drinks packaged in metal cans or plastic bottles, the change in customer’s preferences towards nutritious food and environmentally friendly products created a challenge. To satisfy this new set of customer preferences, PepsiCo took the strategic decision to empower alternatives like Naked Juice and O.N.E Coconut Water (Forbes 2016). The company sourced ingredients from across the world, including certified non-GMO, fresh, and organic alternatives. Further, to enable the production of these products, global supply chains were set-up with refrigeration capabilities throughout the chain. Finally, to satisfy environmentally conscious customers, Naked Juice products were packaged in recyclable bottles. Furthermore, these bottles were explicitly designed in cuboidal shapes to improve packing efficiency during transportation, and the modes of transportation employed were also changed to reduce the overall carbon footprint. Chapters “Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design”, “Supply Chain Risk Management in Indian Manufacturing Industries: An Empirical Study and a Fuzzy Approach”, “Improving Service Supply Chain of Internet Services by Analyzing Online Customer Reviews”, and “An Integrated Problem of Production Scheduling and Transportation in a Two-Stage An Overview of Decisions, Performance and Analytics in Supply Chain Management 5 Supply Chain with Carbon Emission Consideration” address some of the key strategic issues in a supply chain. 2.2 Tactical Tactical decisions, unlike strategic choices, have a relatively moderate impact. Typically, these decisions are related to supply chain planning activities and are made every month or quarter. Due to the shorter time frame, these decisions face lesser uncertainties, though not insignificant. However, companies can leverage better prediction tools to mitigate these medium-term uncertainties and even alter their decisions with relatively more ease. Specifically, strategic decisions taken during the design stage can be capitalized to optimize the supply chain and meet the changing customer demands and market conditions. The tactical decisions required for planning the supply chain activities can be grouped into a few broad categories such as • Purchasing: Decisions on the quantity of materials (supplies) to procure as well as the time to order. • Production planning: Decisions related to the quantities needed to be produced over different time periods to meet varying demands are made. • Transportation: Decisions related to scheduling shipments of raw materials, intermediates, and final products • Inventory Management: Decisions on how much supplies should be stored to mitigate shortage risks while keeping inventory costs minimal. • Distribution: Decisions that aim to coordinate the distributor replenishment schedule with the production capacity to make the product or service available for the customer at the right time IKEA, a multinational furniture retail company, is a success story on how effective tactical decisions can revolutionize a business. IKEA is well-known for providing a wide variety of home furniture at very affordable prices. This is particularly made possible through innovative and optimized decisions in inventory management. For example, IKEA stores observe a “cost-per-touch inventory” principle, where the company seeks to reduce the cost-incurred by cutting down on the number of times it handles (touches) a product (TradeGecko 2018). First, IKEA stores are equipped with showroom inventories from where the customers on selecting their products can retrieve their packages and take them home by themselves. Second, apart from showroom inventory, the store also features reserve inventories of two types—high-flow and low-flow inventories. The highflow inventories are filled with reserve stocks of fast-moving products where more frequent handling takes place. However, by equipping these high-flow facilities with automated storage and retrieval systems, IKEA could further cut-down handling costs efficiently. All the cost-cutting made through the tactical decisions mentioned here is reflected in the low price of the products available for IKEA customers. 6 S. Srinivas et al. Chapters “An Integrated Problem of Production Scheduling and Transportation in a Two-Stage Supply Chain with Carbon Emission Consideration” and “A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for Improving Pharmaceutical Service” prescribe the best course of action for tactical decisionmaking. 2.3 Operational Operational decisions are short-term choices (such as day-to-day operations) characterized by low uncertainty and expenditures. A supply chain is pre-determined by the strategic and tactical decisions, and the operational decisions do not impact its configuration nor planning policies. However, operational decisions can optimize performance at an individual order level within the constraints fixed by the previous decision levels. The focus is to deliver to the inbound customer orders in the most effective manner. Operational decisions are related to activities, including allocating production or inventory in response to customer order, selecting a date for delivering the product or service, updating the pick-up task list used at the warehouse, assigning appropriate shipment methods, and finally, placing replenishment order to maintain inventory. The importance of operational decisions can be emphasized using the case of Target Corporation, one of the largest retailers in the United States. However, the company could not successfully penetrate into the Canadian market. While there are many reasons attributed to the Canadian stores’ failure, ineffective operational decisions played a significant role. These stores ran out of stocks within the initial days of their opening as enough replenishment orders were not being placed. The empty shelves were very disappointing for the eager customers expecting an abundance, as found in the US stores (Fortune 2015). Even during the latter days, the store faced inconsistent supplies at the distribution centers and the retail stores. The flow of individual products was not appropriately tracked, entries were miss-understood, and even the demand forecasts were not accurate. These lead to inefficient operational decisions that were expensive both in terms of storage cost and customer satisfaction. Finally, all the customer resentments cumulated and reached a stage from where recovery seemed far-fetched. Within 2 years, the company shut down all of its 133 Canadian stores and incurred $2.5 billion in losses. Chapters “Pro-Active Strategies in Online Routing” and “Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces” provide strategies to make effective real-time operational decisions. An Overview of Decisions, Performance and Analytics in Supply Chain Management 7 3 Supply Chain Enablers and Drivers Enablers can be understood as things that are needed to achieve a goal. Marien (2000) reports four enablers that are needed for the effective functioning of a supply chain, namely • Organizational infrastructure: It is crucial for an effective supply chain as it determines how the different stages coordinate together to accomplish its goals. The key concern here is organizing supply chain activities, both within the firm and across firms, for vertical orientation or more decentralization. • Technology: Two types of technologies, manufacturing technology, and information technology, are necessary to enable superior supply chains. • Strategic alliances: Establishing long-term partners is a key enabler for a supply chain. Alliances specifically play a vital role in a decentralized supply chain where great power and responsibility are present with the suppliers. • Human resources: Includes technical and managerial employees with a holistic understanding of supply chain management concepts and tools. These employees are needed to design and operate an effective supply chain. While enablers support the smooth functioning of a supply chain, the drivers are areas of crucial decision making. The four major drivers of a supply chain, as described by Ravindran and Warsing (2016, pp. 7–9) and Chopra and Meindl (2013, pp. 41–42), include inventory, transportation, facilities (plants and distribution centers) and suppliers. These drivers do not function independently but interact with each other to establish the overall performance of a supply chain. For example, setting up facilities in remote places away from major cities may reduce rental costs, but will increase transportation costs and affect delivery time. Similarly, procuring and storing supplies in large quantities can reduce cost in terms of raw material and transportation and even improve customer satisfaction levels, but will drastically increase inventory storage costs. Hence understanding how these drivers interact and making efficient trade-offs between them is crucial to achieving superior supply chain performance. 4 Types of Supply Chain Each company has a competitive strategy that involves satisfying the needs of the customer belonging to a particular segment. The overall supply chain and its individual stages need to align with this strategy. For example, customers of a supermarket prioritize availability and variety over the price of the products. These customers are willing to pay higher prices, provided they can buy everything from vegetables to pastries at the same place. Hence, these stores have robust supply 8 S. Srinivas et al. chains that provide a vast range of 15,000 to 60,000 SKUs. While on the other hand, customers of limited-assortment stores want lower prices and are ready to compromise on variety. These stores tend to have cost-effective supply chains that offer a limited range of items (fewer than 2000), with many stores not offering any perishables. The requirements of the customers prioritized by the competitive strategy must match with the capabilities of the supply chain, and this consistency is referred to as strategic fit by Chopra and Meindl (2013, p. 21). In case of a lack of strategic fit, the company must alter the supply chain to meet its competitive strategy or modify its competitive strategy based on what its supply chain is designed to do well. Depending on these supply chain capabilities, supply chains can be understood as one or a mix of the following types: • • • • • • Responsive Efficient Resilient Humanitarian Green Sustainable Figure 2 illustrates the key capabilities of the different supply chain types with respect to seven different criteria, namely, profitability, cost reduction, speed, flexibility, social responsibility, environmental concern, and ethical practice. 4.1 Responsive Supply Chain Ravindran and Warsing (2016, p. 12) describe responsiveness as “the extent to which customer needs and expectations are met, and also the extent to which the supply chain can flexibly accommodate changes in these needs and expectations”. Therefore, responsive supply chains seek to prioritize service levels over operating costs. Further, the responsive supply chains tend to follow a push framework where the supply chain is initiated in anticipation of a customer order instead of a response to an actual customer order (pull strategy). The push strategy helps companies serve their customers quickly but at the cost of higher inventory costs, including wastages. The characteristics of responsive supply chains include: • • • • • Short delivery time Wide product varieties Provision for customized orders High reliability Superior service quality One of the industries that heavily depend on the responsiveness of their supply chains is the fashion industry. To catch up with changing seasonal trends, the companies rely on quick and flexible supply chains. These companies prioritize flexibility, reduced lead time, and timely distribution over cutting costs. For this An Overview of Decisions, Performance and Analytics in Supply Chain Management 9 Fig. 2 Capabilities based on supply chain type with respect to different criteria reason, these companies prefer to have an in-house manufacturing facility despite low-cost manufacturing options in other countries, which might compromise ontime performance or flexibility. 10 S. Srinivas et al. 4.2 Efficient Supply Chain Not every competitive strategy prioritizes responsiveness. In the case of stationery manufacturers, where both the product variety and its demand are mostly standard, responsiveness may not be the key. For instance, adopting a flexible supply chain that manufactures custom stationery in small batches and then ships them using an express transporter only makes the stationery unnecessarily expensive, leaving the customers dissatisfied. In such scenarios, a cost-efficient supply chain is vital. Efficiency is the output obtained per unit input, and in the case of a supply chain, it is the ratio of the revenue generated to the cost incurred. Thus, the sole goal of an efficient supply chain is to minimize the costs. The stages where supply chains focus to cut-costs include: • • • • • Raw materials procurement Inventory holding Manufacturing Transportation and distribution Facility operations Efficient supply chains tend to follow a pull strategy that is practiced in environments where the demand is known. Here the supply chain is only reactive to the actual customer order. Such efficient supply chains tend to hold fewer inventories and carry a level load in warehouses to minimize costs associated with picking and packing. Efficient supply chains have their drawbacks as well. For instance, to reduce costs, product offerings need to be standardized. This affects variety and personalization capabilities, which decreases responsiveness. In fact, for every strategic decision to increase efficiency, there is usually a compromise on responsiveness. This relationship can be observed in the responsiveness-efficiency tradeoff frontier described by Ravindran and Warsing (2016, p. 9–10) and Fisher (1997, pp. 105–117), as shown in Fig. 3. Fig. 3 Responsiveness-efficiency trade-off frontier An Overview of Decisions, Performance and Analytics in Supply Chain Management 11 4.3 Resilient Supply Chain In recent years, a supply chain’s capability to anticipate and handle disruptions (i.e., resilience) has taken importance. This change is primarily due to two reasons. First, today’s world has grown into a highly interconnected global village where a small disruption at a particular place is transmitted to the entire world. For example, earlier, an epidemic was usually contained within a region. However, today, owing to high connectivity, epidemics quickly spread and become a global pandemic like the novel coronavirus disease 2019 (also referred to as COVID-19), causing a worldwide disruption. Secondly, today’s supply chains have become truly global, with each stage located in a different country altogether. Hence a small workers’ strike at one of the countries can stop the entire global supply chain. Thus, in a world of globalization where disruptions are felt everywhere, it is essential to design supply chains by keeping resilience as a priority. In recent times, the COVID-19 global pandemic highlighted the importance of a resilient supply chain. Deloitte (2020) reported that the companies that could perform well during the pandemic were the ones that invested in supply chain risk management. These companies diversified their supply chain from a geographic perspective to avoid risks from disruptions caused in any one country. Further, they multi-sourced vital components to reduce dependency on any one vendor. Finally, they also considered inventory plans that allowed for buffers needed to manage unprecedented disruptions. On the contrary, the companies that scrambled were highly dependent on a specific geography or a particular supplier for vital commodities. A singular focus on cost-cutting caused negligence towards resilience, making these supply chains brittle and vulnerable during disruptions. 4.4 Humanitarian Supply Chain Barve and Yadav (2014) describe the humanitarian supply chain as the “flow of relief aid and the related information between the beneficiaries affected by disaster and the donors so as to minimize human suffering and death”. In a humanitarian supply chain, the customers include the affected people and the intermediate storage facilities, while the supplies include relief aids like materials, logistics, and even volunteers. Since humanitarian supply chains face a lot of unknowns, uncertainties and need coordination among numerous stakeholders (like donors, volunteers, government, NGOs, and military), along with relief aids, the effective flow of information is vital. This type of supply chain is complex as they tend to have limited infrastructure and other resources, making them dependent on donors and volunteers. Humanitarian supply chains need to be highly responsive to the disaster type and its changing phases. While they may be newly established during a particular crisis, humanitarian works during the COVID-19 pandemic showed the importance of 12 S. Srinivas et al. leveraging existing supply chains’ flexibility. Highly responsive fashion companies like Prada, Armani, Gucci, and Giorgio could customize their manufacturing facilities to produce medical overalls. Moreover, within 72 h from the France government’s request, Louis Vuitton, a French luxury conglomerate, converted its perfume factories to manufacture sanitizers and provided pandemic support (Vogue Business 2020). 4.5 Green Supply Chain With increasing customers showing concern for the environment, companies seek to identify and incorporate environmentally friendly practices in their supply chain to gain competitive advantage. Further, the government imposed environmental regulations that have made it imperative for companies to build greener supply chains, especially in countries like France, Spain, Morocco, and Kenya. Building a green supply chain requires a unified effort from all the stakeholders and stages. Manufacturers must work alongside both the suppliers and customers to enable their environmental goals. These environmental goals usually include reducing solid waste, effluent waste, air emissions, and usage of toxic materials. While employing a green supply chain can give a competitive advantage, there still are concerns about whether they will translate into substantial improvement in profits or market share. However, employing environmentally harmful practices have impacted businesses negatively. For example, Nestle, a global food processing company, had to face a myriad of issues when Greenpeace International held the production of Nestle’s confectionery product, KitKat, responsible for deforestation (Purkayastha and Chaudhari 2012). The company was accused of destroying precious rainforests to increase palm plantations and palm oil produce needed to manufacture their confectionary. Further, by facilitating high impact social media campaigns, Green peace could pressurize Nestle to stop sourcing palm oil from Sinar Mars, an organization accused of illegally clearing rainforests, and source its palm oil responsibly. Similarly, Unilever too had to stop purchasing palm oil from controversial vendors. 4.6 Sustainable Supply Chain With dwindling resources, today’s supply chains need to focus on social and environmental facets along with common economic goals to achieve sustained growth. These three dimensions of sustainability are together referred to as the “triple bottom line (3BL)”. The goals of a sustainable supply chain can be understood as the following: An Overview of Decisions, Performance and Analytics in Supply Chain Management 13 • Economic dimension—This is primarily concerned with generating higher profits and achieving growth. • Social dimension—The focus here is on improving employment opportunities, workplace safety, charity, and overall community wellbeing. • Environmental dimension—This deals with aspects involving global warming, ozone layer depletion, climate change, different types of pollution, and ecological preservation. One of the pioneers in sustainability, Ben & Jerry’s, could successfully incorporate all the three dimensions in its mission (Performance Magazine 2020). As early as 1989, the organization opposed the use of recombinant growth hormones to prevent harsh financial impact on family farming. The company introduced the “Caring Dairy” program and established Fair Trade prices to support its farmers in conducting sustainable farming practices. The organization also established Ben & Jerry’s Foundation to motivate its employees to give back to their societies. Further, the company also invested in sustainable packaging solutions. With all its sustainability endeavors, Ben & Jerry’s is considered as an example of how prioritizing sustainability helps businesses build a well-liked brand. 5 Impact of Industry 4.0 on Supply Chain The current industrial revolution, Industry 4.0, is about integrating physical and digital systems to enable effective decisions that require minimal human supervision. It focuses on inter-connectivity using technologies such as the internet of things (IoT), cloud computing, artificial intelligence (AI), and advanced robotics. While these intelligent technologies have transformed numerous areas, they specifically have a profound impact on supply chains. Industry 4.0 has led to major improvements in different supply chain stages, including procurement, inventory management, logistics, production, and retailing, by facilitating process integration, automation, digitization, and analytical power. The ability to exchange data and make decisions is particularly useful in supply chain as they have a dynamic network consisting of multiple stakeholders and stages needing collaborative decisions at every level. Despite numerous benefits, a majority of companies are yet to adopt Industry 4.0 technologies into their supply chain due to doubts on return on investments, struggle to find qualified talent to implement and maintain these systems, shortage of financial resources, concerns over data security, lack of information technology infrastructure or even due to the sheer lack of knowledge about its benefits (Horváth and Szabó 2019). However, there have indeed been numerous success stories. One of the best examples is the widespread adoption of intelligent tools that can accurately predict consumer behavior and guide demand planners. Demand planning is a manually intensive week-long task repeated at the beginning of every month (SAS 2018). Demand planners spend 40% of their time cleaning and managing inventory and sales data, an additional 30-40% of time reviewing forecasting models and fine- 14 S. Srinivas et al. tuning them, and 10% of time reporting their findings (SAS 2018). This further becomes cumbersome when a supply chain deals with numerous products across different categories. Nevertheless, with developments in analytical technologies, supply chains leverage AI tools to effectively handle most of these manual and repetitive tasks, leaving time for demand planners to focus on other value-adding work. Apart from demand planning and then sourcing from reputed suppliers, ensuring the quality of a product/service is a vital aspect of supply chain management (Romano and Vinelli 2001). With improvements in capturing and analyzing sensory data, companies continuously monitor their product/service to ensure consistent quality. For example, manufacturing industries capitalize on image processing and machine learning developments to automate visual inspections, which would otherwise be time-consuming, labor-intensive, and prone to human error. Like few modern call centers, other service providers also leverage voice recognition and natural language processing tools to ensure their agents’ quality and drive customer satisfaction. Further, with developments in IoT technologies, not only quality is monitored but also controlled. For example, many plastic goods manufacturing companies continuously monitor their manufacturing process at all stages. When any temperature deviation or product abnormality is identified, the production process enabled using IoT is automatically modified to produce ideal results. Thus, Industry 4.0 also supports supply chains by enabling consistent quality, cuttingdown costs, improving resource utilization, and driving process efficiency. Due to the developments in Industry 4.0 technologies, specifically IoT and sensor technology, both manufacturing and service organizations generate massive amounts of industrial data. To a large extent, the success of a supply chain lies in its ability to effectively capitalize on this data by leveraging advanced analytical tools to gain intelligence. Supply chain analytics deals with the use of quantitative models for data-driven management of all the decision levels—from helping in supply network design and vendor selection at the strategic level to managing procurement, inventory, demand planning, and logistics at tactical and operational levels. The traditional sources of data for supply chain analytics include radio frequency identification (RFID) systems that automatically track items attached with RFID tags, global positioning systems (GPS) that provide the locations of shipments in transit, and barcode enabled systems that capture transactions. Further, supply chain analytics also heavily depends on data visualization techniques to report its findings in a human interpretable manner. Depending on the complexity and value addition, supply chain analytics can be classified into three types—descriptive, predictive, and prescriptive. • Descriptive analytics uses aggregation, visualization, and mining of historical data to provide insights on prior trends and patterns. It helps supply chain practitioners to learn from the past and also uncover relationships between variables, thus empowering them to make better decisions in the future. Apart from past data, descriptive analytics also uses current data to provide valuable real-time information and visibility needed to effectively manage the supply An Overview of Decisions, Performance and Analytics in Supply Chain Management 15 chain. For example, information on the current locations and quantities of different products in a supply chain can help managers optimize several decisions such as delivery schedule, transportation modes, and replenishment orders. Tools such as dashboards, scorecards, and sales reports are key enablers of descriptive analytics. • Predictive analytics enables an organization to estimate the likelihood of future outcomes using historical data. It relies on forecasting and machine learning algorithms to achieve its purpose. In a supply chain, predictive analytics is majorly used to predict demand, customer purchasing patterns and behaviors, inventory records, and performances of various supply chain stages. The complexity of the models deployed as well as the value-addition from predictive analytics is higher as opposed to descriptive analytics. • Prescriptive analytics aims to provide the best course of action, given a particular situation, and also report on the consequence of undertaking such an action. While descriptive and predictive analytics provide decision support, prescriptive analytics ventures one-step further with decision automation. For example, predictive analytics can forecast the demand given historical data, whereas prescriptive analytics will be able to capitalize on such predictions to provide the optimal replenishment policy along with its impact on the inventory costs. Thus, prescriptive analytics provides the highest degree of intelligence, but is also the most complicated among the three types of analytics. The key tools for prescriptive analytics are mathematical optimization models, simulations, and heuristics. Though many companies have employed prescriptive analytics to automatically optimize production, inventory, and logistics, the use of prescriptive analytics is still at its early stages (Lepenioti et al. 2020) This book focuses on introducing the recent trends in supply chain management as well as the applications of advanced analytical models that impact different decision levels. While adoption of Industry technologies in supply chains has numerous facets, improving visibility is one of them. Visibility provides the ability to track orders and products as they move through the manufactures’ value chain to the final customer. Chapter “Intelligent Digital Supply Chains” provides a detailed discussion on visibility in supply chains. Long-term decisions need to include future uncertainties, and Chaps. “Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design” and “Supply Chain Risk Management in Indian Manufacturing Industries: An Empirical Study and a Fuzzy Approach” strive to optimize these strategic decisions in the manufacturing sector by proposing a mixedinteger linear programming model and multi-criteria decision-making methods, respectively. Particularly, Chap. “Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design” is concerned with strategic decisions needed to optimize the product life cycle in a specialized supply chain, while Chap. “Supply Chain Risk Management in Indian Manufacturing Industries: An Empirical Study and a Fuzzy Approach” aims to address strategic decisions that aid in foreseeing exigency and mitigating risks. Chapter “Improving Service Supply Chain of Internet Services by Analyzing Online Customer Reviews” focuses on the service 16 S. Srinivas et al. supply chain, specifically that of internet services. It aims to discover strategic insights from online customer reviews by employing text analytics and root cause analysis techniques for this purpose. To address the need of the hour, environmentfriendly practices, Chap. “An Integrated Problem of Production Scheduling and Transportation in a Two-Stage Supply Chain with Carbon Emission Consideration” proposes a mathematical model for integrated optimization of strategic and tactical decisions that also accounts for carbon emissions. Chapter “A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for Improving Pharmaceutical Service” addresses tactical decisions in service delivery using simulation modeling. It tries to evaluate the integration of futuristic drones for delivering vital pharmaceutical products. Finally, Chaps. “Pro-active Strategies in Online Routing” and “Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces” propose optimization models and heuristics to deal with operational decisions proactively. While Chap. “Pro-active Strategies in Online Routing” aims to control urgent logistics in real-time, Chap. “Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces” focuses on prescriptive analytics for controlling scheduling in a manufacturing process in real-time. References Barve, A., & Yadav, D. K. (2014). Developing a framework to study the various issues of humanitarian supply chains. In Proceedings of 10th Asian Business Research Conference– Bangkok, Thailand. Blackstone, J. H. (2010). APICS dictionary, APICS the Association for Operations Management (13th ed. revised, p. 148). Chopra, S., & Meindl, P. (2013). Supply chain management: Strategy, planning, and operation (5th ed.). London: Pearson Education. Deloitte. (2020). COVID-19 - Managing supply chain risk and disruption. Fisher, M. L. (1997). What is the right supply chain for your product? Harvard Business Review, 75, 105–117. Forbes. (2016). PepsiCo’s practical application of supply chain resilience strategies. Retrieved from https://www.forbes.com/sites/stevebanker/2016/10/01/pepsicos-practical-application-ofsupply-chain-resilience-strategies/#1d81bfea6293 Fortune. (2015). Why target failed in Canada. Retrieved from https://fortune.com/2015/01/15/ target-canada-fail/ Horváth, D., & Szabó, R. Z. (2019). Driving forces and barriers of Industry 4.0: Do multinational and small and medium-sized companies have equal opportunities? Technological Forecasting and Social Change, 146, 119–132. Lepenioti, K., Bousdekis, A., Apostolou, D., & Mentzas, G. (2020). Prescriptive analytics: Literature review and research challenges. International Journal of Information Management, 50, 57–70. Marien, E. J. (2000). The four supply chain enablers. Supply Chain Management Review, 4(1), 60–68. An Overview of Decisions, Performance and Analytics in Supply Chain Management 17 Performance Magazine. (2020). Ben & Jerry’s – An example of how to integrate sustainability in business. Retrieved from https://www.performancemagazine.org/ben-jerrys-sustainabilitybusiness/ Purkayastha, D., & Chaudhari, A. (2012). Greenpeace, nestle and palm oil controversy: Social media driving change? IBS Center for Management Research. Ravindran, A. R., & Warsing, D., Jr. (2016). Supply chain engineering: Models and applications. Boca Raton, FL: CRC Press. Romano, P., & Vinelli, A. (2001). Quality management in a supply chain perspective. International Journal of Operations & Production Management, 21(4), 446–460. SAS Institute. (2018). Assisted demand planning using machine learning for CPG and retail. TradeGecko. (2018). IKEA supply chain: How does IKEA manage its inventory? Retrieved from https://www.tradegecko.com/blog/supply-chain-management/ikeas-inventory-managementstrategy-ikea Vogue Business. (2020). Luxury’s war effort against coronavirus. Retrieved from https:// www.voguebusiness.com/companies/european-luxury-fashions-war-effort-against-coronavirus Intelligent Digital Supply Chains Raghav Jandhyala 1 Introduction Supply chains are growing increasingly complex with global, multi-enterprise and multi-channel operations to meet the ever-growing customer expectations and need for individualized products. In this intelligent era, supply chains need to be digitized in order to be responsive, sustainable, profitable and resilient to disruptions. The flow of information across all aspects of the supply chain from strategy to execution and the collaboration across inter- enterprises and cross functional units are a necessity of modern supply chains. An end-to-end visibility of connected supply chain data is key to analyze, sense and respond to changing market situations. In this chapter, we will cover the digitization of supply chains and intelligent visibility with a focus on the following topics: • • • • • • • • • • • Digital supply chains. Intelligent visibility connecting strategic, tactical and operational processes. Supply chain control tower—global and local. Next-gen supply chain analytics. Supply chain alerts and exception management. Insight-to-action on supply chain issues. Supply chain performance metrics—KPIs. Cognitive supply chains—role of technology in Industry 4.0. Resilient supply chains. Evaluation of supply chain assumptions, risks and opportunities. Collaborative enterprise planning integrating finance with operations. R. Jandhyala () SAP Labs, Tempe, AZ, USA e-mail: raghavjv@gmail.com © Springer Nature Switzerland AG 2021 S. Srinivas et al. (eds.), Supply Chain Management in Manufacturing and Service Systems, International Series in Operations Research & Management Science 304, https://doi.org/10.1007/978-3-030-69265-0_2 19 20 R. Jandhyala 2 Digital Supply Chains Digital supply chains are the next generation real time supply chains with seamless visibility of data and well-orchestrated planning, execution, cross-collaboration, analytics and intelligence and automation across all operations of a supply chain. Digital supply chains enable supply chain business processes such as Designto-Operate. Before we get into the digital supply chain, let us first understand the challenges and opportunities in a supply chain network. 2.1 Challenges in Digital Supply Chain Silos of information across different functional units and operational boundaries with disjointed strategic, operational and execution plans along with disconnected manufacturing, operations and logistics make it harder for companies to respond to customer demand shifts. Below are some of the most common challenges faced by supply chains. 1. Lack of visibility into the health of the supply chain, inventory position across different locations, in-transit inventory and manufacturing bottlenecks. 2. Disconnected systems involved in supply chains needing manual steps for integration along with multiple points of failures. 3. Lost sales because of inability to get a full demand and supply picture and run simulations. 4. Disconnected planning and execution processes and systems. 5. Lack of collaboration across stakeholders involving sales, marketing, finance, and supply chain with many ad-hoc communications and decisions lost in emails. 6. Offline disconnected excel spreadsheets to store and plan supply chains with complex hidden rules. 7. Lack of what-if scenario planning where planners cannot wait for information to be available the next day or next week from separate systems for planning, execution, reporting and analysis. The ecosystem of a supply chain in the modern world is complex with customer networks, supplier networks, logistic networks, contract manufacturers, in-house network and inter/intra company routing policies. Further, modern supply chains are complex, owing to global manufacturing, inter-continental distribution network, and mergers and acquisitions (Kepczynski et al. 2019a, 2019b). Intelligent Digital Supply Chains 21 2.2 Business Processes Evolution and Trends in Digital Supply Chain With the changes in supply chain processes, thin margin and availability of technology, business processes have changed over time, moving from a traditional waterfall planning and execution to a more continuous planning and execution. A closed loop and continuous alignment of planning and execution take the business to a new level, which was not possible before, connecting the strategic and tactical processes to the operations/execution and responding to short term disruptions along with incorporating better customer and product experiences. Integrated Processes Supply chains have evolved from being fragmented with departmental planning and execution to a collaborative one with integrated planning across demand, supply and finance. Faster Planning Cycles By bringing data to one system of planning and execution, customers now have a greater visibility of data and can plan faster and frequently compared to the traditional monthly or quarterly planning processes. Accuracy of Supply Chain Models With the advances in forecasting and planning algorithms across demand, supply and inventory, enterprises have shifted from being supply driven to demand driven by able to sense customer demand and respond to changes more efficiently. Improved Experiences With better visibility into customer and product experience, enterprises can support mass manufacturing of products and bring more customization choices to their end customers. Figure 1 shows an evolution of supply chain planning from efficiency to experience. This figure is adopted from SAP positioning of digital supply chain. Though the evolution of supply chains over the last few decades has dramatically increased, the next generation, state of the art supply chains are moving towards bringing: Autonomous Self-Regulated Capabilities Supply chains that consider historical trends, internal and external events, and signals to provide recommendations, based on corporate goals assist the planners with informed decisions and supply chain executives with strategies to drive the business forward. Real time Synchronized Planning The traditional planning with monthly and weekly cadence will evolve towards continuous and real time planning. This will help to think beyond the traditional demand and supply streams to a new valuechain oriented supply chain view across the end to end supply chain. Network Collaboration The business processes collaboration and visibility are increasingly getting better within an enterprise. However, as seen in the recent COVID-19 pandemic, supply chains are global, and business processes are interconnected with enterprises beyond their own chains. A disruption in one of the 22 R. Jandhyala Fig. 1 Evolution of supply chain planning enterprises has an impact on the end to end supply chain. Therefore, a network of inter-enterprises with collaboration and visibility into business processes, enabling easy onboarding of business partners into a truly scalable network, is required. 2.3 New Business Models Enabled by Digital Supply Chain By digitizing the supply chain, companies can expand beyond their traditional boundaries and enter new markets with new business models. With the increasing competition and continuously changing customer behavior, companies need to evolve to be innovative and survive the market disruptions. Figure 2 shows the new business models enabled by the digital supply chain, as adopted from SAP’s positioning of Digital Supply Chain of One. Consumer Buying Patterns Consumers in the current digital world have multiple avenues to shop for the best product and get the best competitive price. Customers have different buying patterns. They can even search online and buy at store, or physically view that product in a store and buy online. Customers’ buying decisions are influenced based on product reviews and trends on social media. The physical and virtual boundaries are blurring for consumers, and the companies need to be innovative to meet or exceed consumers’ expectations. Examples of digital disruption by innovative companies: Uber is the largest taxi company without owning any vehicle. Netflix is the largest streaming company without owning movies; however, they later changed their business model to launch Intelligent Digital Supply Chains 23 Fig. 2 New business models enabled by digital supply chain Netflix movies. Airbnb is the largest accommodation provider without owning any property. Wholesale to Consumer Model Consumer product companies are expanding their business boundaries to open online channels to reach consumers directly; they now ship their products directly to the consumers in addition to supplying them to retailers. This also helps them to sense and respond directly to the changing market or customer buying patterns and serve better products to the customer. These companies now have a better pulse on the consumers as they do not solely rely on the retailers to provide the demand signals. Retail Supply Chains Retailers are entering the consumer product market by providing their own brands and products at a competitive price. Retailers have a better sense of the customer demand and buying patterns by analyzing the point-of-sales transactions (POS) and are expanding their supply chain footprints with their own distribution centers and manufacturing plants for products that can bring additional revenue. For example, Safeway provides signature branded products across different departmental units. Walmart offers “Great Value” brand for various products manufactured by Walmart. Retailers to Logistic Providers With customer demand ever-increasing to get the product physically within the same day or the same hour, we see retail companies entering the logistic business. Online retailers like Amazon are entering the logistic space with their own fleet of logistics or a mix of inhouse or external logistics partners. Companies like Albertsons own and partner with logistic providers for same-day delivery of items purchased online. Best Buy and Home Depot offer in-store pickup and delivery services. Companies that offer a high level of logistics flexibility see increasing loyalty with customers and now have firsthand information on customer demand and better visibility of their inventory position. 24 R. Jandhyala Logistics to Manufacturers Logistic companies, to be competitive and provide the best customer experience, are investing in 3D manufacturing to reduce the costs and carbon impact of transportation wherever it is feasible. For example, auto logistic companies for spare parts are partnering or establishing 3D manufacturing sites for spare parts in the hubs where there is more customer demand, to reduce interstate transportation and optimize intercity delivery. Manufacturers to Service Delivery With the digital supply chain and online digital twin representation of a physical asset, manufacturers can get information on the asset usage and proactively provide services for maintenance to the consumer. This integrated manufacturing and servicing model generates additional revenue streams, better customer service, and reduces costs. 2.4 Design-to-Operate Business Process Design-to-Operate is a supply chain centric business process that enables companies to digitally connect the end-to-end supply chain, from designing new products and assets to managing them throughout the lifecycle from planning to manufacturing, delivery and operation in the field. Such a business process delivers the speed of innovation, operational efficiency and service effectiveness necessary to meet and satisfy customer expectations. Figure 3 shows an end-to-end business flow of Design-To-Operate scenario for digital supply chain as adapted from SAP’s positioning of design-to-operate. It includes design, plan, manufacture, delivery, and operations aspects across the connected business network of customers, suppliers, manufacturers, assets and logistics. Example of Design-To- Operate process for an Automobile customer 1. Design: The design process allows the supply chain specialist to design the next generation car based on market trends, customer sentiments, and technological advancements. For example, the market shows a trend in buying electric and diesel cars. Supply chain designers can create several options for the car like 2 door vs. 4 door, different desired colors e.g. blue matte, or an electric range for the car. Supply chain designers can also validate their design with a focus group of customers to perform a conjoint analysis and get their feedback on different variants of the cars and the price they are willing to pay for different feature options. 2. Plan: Supply chain planners can then create a feasible plan considering customer demand, sales history, macro-economic factors, and customer sentiments to arrive at a future forecast. A supply plan is then created for the demand by performing capacity and materials planning with decisions to make, buy from Intelligent Digital Supply Chains 25 Fig. 3 Design-to-operate process details third part or outsource parts of the manufacturing process. A financially viable plan is created to meet the customer demand. 3. Manufacture: In the manufacturing phase, the different phases of manufacturing are carried out by collaborating and procuring raw materials from suppliers, assembling semi-finished goods, and scheduling, in detail, the production of finished goods, whether in-house or through contract manufacturers. Further, a digital twin of the car is created and stored as a virtual asset to enable monitoring by several cross-functional units. 4. Deliver: Finished goods, in this case, cars, are delivered to the customers based on the customer orders received. To meet the customer service levels for on-time delivery, the supply chain needs to be digitally connected to the logistic network to determine the optimal routing, warehouse capacity, and disruptions based on internal and external events. 5. Operate: The asset received by the customer is continuously monitored, and key performance indicators (KPIs) like mileage and usage are sent to the digital twin. The IoT sensors in the asset continuously analyze the asset and predict any failures. This monitored data is continuously used to build better products by the manufacturer. Further, if there is a need for maintenance, the customer is alerted, and the asset is scheduled for a service through proactive and predictive maintenance of the asset. 3 Intelligent Visibility in Supply Chain Networks Organizations need to have a unified view of supply chain plans and the execution of such plans to enable a real time supply chain that profitably balances demand and supply. Organizations need to react faster to changes and break departmental silos of information by enabling a connected supply chain that can reduce planning cycle 26 R. Jandhyala time and run scenarios for decision making. An intelligent supply chain enables organizations to profitably analyze, sense and react to changes in real time to meet the supply chain goals of • • • • • Increasing supply chain agility Increasing end-to-end visibility Reducing cycle time for planning Increasing on-time delivery and customer service levels Reducing supply chain cost In order to realize the end to end digital supply chain, it is important to have an end to end visibility of all the processes and activities across several stakeholders and interconnected networks. These include: • • • • Visibility of planning processes for strategic, tactical and operational planning Close loop alignment of planning and execution KPIs to measure supply chain performance and health Real time visibility of events that happen during execution. Figure 4 shows the systems, people, processes and tools across a supply chain. These include planning types ranging from strategic to execution, business processes covering sales and operations planning to manufacturing and sales & distribution, time horizons ranging from years to hours, dimensionality ranging from aggregated business units to detailed order level, systems ranging from planning to logistics systems, and supply chain roles ranging from sales and marketing to operations. The subsequent chapters will cover some details of each of these areas that make an end to end supply chain visibility challenging and the benefits of having a truly integrated supply chain system. Fig. 4 Supply chain systems, people, process and tools Intelligent Digital Supply Chains 27 3.1 Challenges Achieving End-to-End Visibility • Multiple systems are involved in a supply chain process with disconnected data and processes • Data warehouses with a lack of visibility across all involved parties. Misaligned data and rigid models. • Weekly/nightly processing of data to warehouses that delays visibility into problems • No proper visibility related to tracking of adherence to the plan. • Disconnected functional units and processes—siloed and stale data. • No visibility outside of the four walls of the supply chain network. e.g., lack of knowledge on whether the supplier has enough inventory, and if they can commit to the forecasts to keep up customer service levels. • Lack of visibility to demand, supply, inventory along with insights into strategic plans, execution, and financial targets. 3.2 Benefits of Having Visibility in Digital Supply Chain Include • Reduction in supply chain costs • Improved customer service levels by proactively addressing lost sales and demand shortages • Reduced obsolescence or excess inventory • Increased productivity without firefighting • Improved revenue and profit margins 4 Global and Local Control Towers Providing Global E2E Visibility A supply chain control tower is a system of record for supply chain visibility that connects the data across multiple sources to provide an end-to-end visibility, performance management, exception management, case resolution, and simulation capabilities. It provides insights into the current problems in the supply chain and events that result in plan deviations, and it projects the future outcome by mitigating the risks. There is an increasing demand in the market for supply chain control towers that provide real time insights into the internal and external events which affect the normal operations of the supply chain and reduce the time spent by users to identify, analyze and resolve issues in a timely manner. With the advances in technologies and artificial intelligence, there is a need for autonomous supply chain control 28 R. Jandhyala Fig. 5 Global and local supply chain control towers towers that can self-identify and heal common and repeating pattern issues. Supply chain control towers are not the systems for planning or execution, but they provide visibility across the end to end supply chain with a single source of truth for decision making. For example, a supply chain control tower can provide visibility that a sales order will not be delivered on time and will affect the order fulfillment KPIs with the strategic customer. The resolution for this issue is typically performed in the planning or execution system. Supply chain control towers can be classified into the global control tower and local control tower. Global control tower brings focus on key aspects that affect the movement of materials across the supply chain with visibility to inbounds, outbounds and in-transits across the internal and external networks. A global control tower may provide visibility to critical events, and the details of such events are managed in a local control tower. For example, a local control tower for manufacturing will have visibility over the production plans, deployments and detailed scheduling, whereas the control tower for logistics will have information about freight orders, transportation schedules and other execution data. Once the data is sourced from multiple systems, it is important to identify which data is relevant for visibility and only bring that data to a global control tower and manage details in a local control tower. A global control tower would primarily have insights into the overall demand, supply and inventory positions along with insights into execution data related to the movement of goods. Further, any events from the local control tower, for example, a truck delay should be visible in the global control tower for visibility into the issues and decision making. Figure 5 shows local control towers of sales, assets, manufacturing and logistics providing insights to the global control tower for endto-end visibility, scenario planning and collaboration. This figure is adapted from SAP’s positioning of Integrated Business Planning (IBP) Supply Chain Control Tower. Intelligent Digital Supply Chains 29 Supply chain control towers also connect digitally with the extended supply chain networks including suppliers, contract manufacturers and logistic providers. For example, a forecast commit scenario between manufacturer and supplier digitally connects the requirements of raw materials for manufacturing of finished goods by the buyer with the forecast commit of raw materials from the supplier. Further visibility of supplier’s inventory allows the buyer to make better decisions on which customer demands to fulfill. This reduces uncertainty in planning and better transparency across all stakeholders of the extended supply chain. Supply Chain Systems of Record Most enterprises have multiple supply chain systems to manage the end to end supply chain, each with different purpose. Thus, for an organization, a supply chain is an interconnected system of systems. Some of the common systems in the supply chain are listed below: • One or more enterprise resource planning (ERP) systems as execution system of records based on different geographic regions or different business units. Examples include SAP ECC, SAP S4HANA and JDA. • One or more planning systems as planning system of record based on different geographic regions or different business units. Examples include SAP Integrated Business Planning, JDA, Kinaxis, and E2OPEN. • External systems for collaboration with suppliers and customers. Examples include SAP ARIBA and Oracle. • External pricing systems. • Warehouse management systems. • Transportation management systems. • Logistic business network providers. • Systems from mergers and acquisitions. • Experience management systems for unstructured data/customer sentiment analysis. • Data warehouses for standard reporting and KPIs. • 3PL third party systems. Although it would be desirable to have a single system which can manage the end to end supply chain, in reality, the enterprises have multiple supply chain systems. Visibility to data across all these systems gives insights into supply chain health to identify potential issues and take corrective measures. A supply chain control tower connects the data across these multiple sources to provide an end-to-end visibility and single source of truth across all stakeholders. It provides instant insights into the demand, inventory and supply across the entire supply chain with visualizations in the form of supply chain network diagram, geographic views, heatmaps and charts. However, it should be carefully managed as to what data should be included for visibility because bringing data across all these systems into a single system for visibility would be a massive amount of records with high latency, multiple points of failure, storage costs, and noise which defeats the purpose of gaining useful 30 R. Jandhyala Supply Chain Control Tower integrates performance data from multiple systems SAP Supply Chain Control Tower SAP Integrated Business Planning (SAP IBP) SNC ERP ERP S/4 HANA S/4 HANA APO 3rd Party ERP S/4 HANA Inventory Management Manufacturing Insights 3rd Party 3rd Party Fig. 6 Multiple supply chain system of records insights from data. Therefore, it is important for end-to-end visibility to bring in the key information from these different systems with aligned master data and business semantics and present to the end users at the right level of granularity for analysis and decision making. Figure 6 is adapted from SAP’s positioning of IBP Supply Chain Control Tower and shows an example of several supply chain systems across different regions and business units. These include systems like SAP ERP or SAP S/4 HANA for supply chain execution, Advanced Planning and Optimization (APO) for supply chain planning, SAP Integrated Business Planning (IBP) for next-gen supply chain planning, SAP SNC and SAP ARIBA for collaboration with external suppliers and several other third party systems that are integrated to SAP IBP Supply Chain Control Tower. 5 Next-Generation Supply Chain Analytics Analytics on real-time supply chain data is very much needed to understand the health of the supply chain and make data-driven decisions. Analytics has been a very important and integral part of supply chain management, especially needed to visualize and make a meaningful analysis of the vast amount of supply chain data. Analytics provides a graphical way to understand the supply chain data, trends, seasonal and buying patterns, comparisons (cost-to-serve by region), and view anomalies and outliers (e.g., promotions in sales history) in the data. Analytics provides end to end visibility of inbound and outbound activities that happen in a supply chain by integrating important information across multiple source systems. It Intelligent Digital Supply Chains 31 is the basis for managing exceptions and driving actionable results from the derived insights in supply chain networks. Supply Chain Data Is a Big Data Problem An enterprise with 1000 customers, 100 products, 10 distribution centers, and a few factories can plan millions of combinations of products, customers and locations. Add to that the time series data of past 2 years and predictions on next 3 years in weekly granularity and multiple plans across sales, marketing, supply chain and finance. This can easily result in billions of planning data points. As supply chain data grows and changes rapidly with accumulating sales, production, procurement, transportation and events data, it is necessary to view the complete picture of a supply chain rather than a narrow focus on a particular business process or business unit or region. Even with technological advancements, there are several challenges that organizations face to get decent analytics 1. No real time visibility of data because of long ETL (Extract, Transform and Load) processes that bring data to an enterprise data warehouse from transactional systems. This can take sometimes days or weeks for users to see the latest and accurate data in the enterprise data warehouse. 2. Multiple supply chain systems of data (ERP, planning, logistics, warehouse, etc.) with a lack of integration and semantics matching across systems. 3. IT overhead to create and maintain analytic charts and reports. 4. Fixed analytics with a lack of flexibility to create different chart types or perform drill-downs or filter data. 5. Offline spreadsheets and charts to multiple copies of data that get stale over time. 6. No single version of truth resulting in poor management review meetings where the discussions and firefighting are on who has the right data rather than making decisions based on one version of the truth. Beyond Visualization The technology advances in storing and analyzing data have opened new avenues for supply chain executives to view one version of the truth and make data driven decisions based on real time data. Such an integrated analytics system allows supply chain analysts to go beyond simple visualizations of charts to a comprehensive analysis and decision-making process, which includes: 1. 2. 3. 4. Aligned supply chain KPIs and metrics. Balances scorecard of key supply chain functions. Transparency of information and collaboration across various functional units. Exception management with insights: Automated alerts along with insights into the root cause and solution recommendations. 5. Predictions of events that disrupt to supply chain. For example, container shipping late. 6. Real time ‘what if’ analysis and scenario planning to evaluate several solutions along with the financial impact of the decisions. 7. Connected supply chain sources with an end to end visibility of the network, both internal and external. 32 R. Jandhyala Fig. 7 Supply chain analytics dashboard 8. Navigation to the source system of records e.g., sales orders in ERP to make changes to data. Figure 7 shows an example of a supply chain dashboard with the following chart types • • • • • • Geo charts showing manufacturing locations Heatmap presenting capacity utilization Horizontal bar chart indicating resource overload Pie chart showing sales by country Combination line and bar chart comparing financial plan and forecast projections Vertical bar chart reporting actual revenue by products segmented to A, B and C categories. 5.1 Common Analytical Charts Relevant for Supply Chain Analytical charts can be classified for different functional areas of supply chain e.g., demand, supply, inventory, financial reconciliation, or across process types from strategy, tactical to operational plan across different time horizons, organizational structures and roles (Kepczynski et al. 2019a, 2019b; Kusters et al. 2018; Chopra et al. 2013). Figures 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, and 21 are some of the common charts in a supply chain as adapted from SAP Integrated Business Planning. 33 Fig. 8 Forecast analysis Intelligent Digital Supply Chains 34 R. Jandhyala Fig. 9 Customer service levels Fig. 10 Time series characteristics Fig. 11 Forecast error and bias 1. Forecast analysis: Line chart of forecast vs. actuals comparing statistical forecast generated from sales history, sales quantity and actuals quantity Intelligent Digital Supply Chains 35 Fig. 12 Plan adherence to financial target Fig. 13 Inventory types valuation 2. Customer service levels: service levels by product family, customer segment and location with drill down options to view customer fill rates. 3. Time series analysis: Distribution of time series property across the demand inputs. This shows a pie chart of data distributed by continuous, seasonal and trend data. 4. Forecast error and bias: A vertical bar chart by time series comparing the forecast error and bias in the planned forecast vs. actuals for a 3 month lag. 5. Adherence to financial targets: compare consensus plans to the annual operating plans and actuals quantity projection from the prior year across all product families. 6. Inventory types valuation: inventory valuation for the next 12 weeks distributed across safety stock, cycle stock, merchandise stock, pipeline, and in-transit inventory. 36 R. Jandhyala Fig. 14 Global inventory position By Material Type ($USD) 25M Projected Inventory Target Inventory Shortage 20M 15M 10M 5M 0 Finished Goods Semi-Finished Raw-Materials Packaging Fig. 15 Inventory projection by material type 7. Global inventory position: geo chart of stocking nodes in the supply chain showing the on-hand position of inventory and locations with excess inventory and low inventory. 8. Inventory projections by material types: view inventory position by finished goods, raw materials, semi-finished goods and packing materials comparing against inventory targets. Provides visibility into projected inventory shortages below safety stock. 9. Production and distribution network: gives visibility to material flows, in-transit quantities across the network, late shipments, etc. Figure 16 shows a supply chain network visualization of all the downstream and upstream nodes of the network from customers to distribution centers to manufacturing plants to suppliers along with the production process from raw materials to finished goods. Intelligent Digital Supply Chains 37 Fig. 16 Production and distribution network 250 200 150 100 50 0 Sim - Cap Usage PR Fig. 17 Capacity overheads Fig. 18 Customer demand fulfillment: demand vs. supply 10. Resource capacity views: shows the available capacity vs. the capacity usage of the products. It provides visibility into resources that are over or under-utilized with drill down capabilities to see which products are consuming more capacity. 38 R. Jandhyala Fig. 19 Process monitoring Fig. 20 KPI micro charts Projected Inventory vs Demand 800K 600K 400K 200K W1 W2 W3 W4 W5 W6 W7 W8 Fig. 21 Projected Inventory and Demand 11. Customer demand fulfillment: includes a horizontal bar chart of customer demand vs. constrained demand for the product families considering supply constraints 12. Process monitoring and collaboration: a process chart showing the current state of the planning process, its due dates, tasks, and people collaborating in the activity. 13. Key performance Indicators: these include KPI charts for performance metrics. E.g., plan adherence, customer service levels, days of coverage and out of stock percentages. Intelligent Digital Supply Chains 39 14. Complete demand, supply and inventory: include charts showing both volumes and financial valuation of quantities allowing drill downs into demand (forecast, sales orders, dependent demand), inventory (current and projected inventory by finished goods, raw, pack and work-in-progress) and supply elements (distribution receipts, in-transit inventory, production planned, production confirmed) 5.2 Business User Friendly Self-Service Analytics Modern supply chain systems should provide business user friendly analytics capabilities that help the user to quickly get to the right information without heavily relying on the IT team to provide the necessary reports, data and models. Usability of analytics and good visualizations are critical for the adoption of analytics by the end user community. Below are some of the characters of analytical tools which help users unleash the business value of data. • Pre-delivered content with the standard set of dashboards, charts and reports for most common areas in supply chain • Self-service creation of charts specific to the user’s needs with very minimal training. • Multiple visualizations to represent the data in the right format which could be both table and chart format. • Various charting options like network charts, geographic chart, pie charts, bar chart, dual axis charts and waterfall charts. • Define and visualize thresholds for the key figure measures on the charts • Easy filtering options on the data represented in the charts by time ranges or supply chain dimensions like products, customers, locations, resources, etc. • Drill down into the details of a specific section of the chart allowing users to narrow down on an issue. • Flexible slice and dice of data across different hierarchy levels. • Easily define and view top N and bottom N values. • Smart insights into the information beyond what is displayed on the chart • Annotations on different sections of the chart • Faster performance on charts with faster refreshes of data. • Collaboration of analytics with other stakeholders by sharing charts. 6 Supply Chain Alerts and Exception Management Alerts or exceptions to the planned results occur very frequently in supply chains. In order to focus attention on the deviations in the supply chain, users require robust alert management functions that work on the vast amount of data and provide manageable alerts at the level which can be understood acted upon. 40 R. Jandhyala Examples of alerts include: • Sales orders with late delivery. • Capacity overload exceptions where resource capacity usage is projected over 100%. • Forecast accuracy of <80% for a product group. • Projected inventory over target inventory for consecutive periods. • Demand shortage. • Customer and vendor order fulfilment alerts. • In-transits overdue. • Production overdue. • Raw material shortages. • Demand fluctuations. • Supply variability. Figure 22 shows an example of an alerts overview generated across different alert categories in a supply chain: capacity overloads, forecast below thresholds, orders vs. forecast for various product segments. Such alert overview charts provide the number of alerts generated in each category along with a priority label of high, medium or low such that the user can focus on the important alerts and drill down into its details. This figure and others in this section are adapted from SAP Integrated Business Planning to illustrate the flexible configuration and usage of alerts. An alert framework should be flexible to allow business users to create their own alert conditions. The alerts generated are real time on the fly based on the alert definition. The main components of alert definitions include: Alert description: qualitative information about the alert. e.g., capacity overload alert. Alert calculation levels: the aggregated levels at which the alert results are calculated by rolling up the data across dimensions and time, e.g., weekly, resource and location level. Fig. 22 Alerts overview Intelligent Digital Supply Chains 41 Time horizon: the level and period range for the alert to be calculated, e.g., 2020 calendar week 20 to 2021 calendar week 19. Severity: the priority of alerts: high, medium, or low. Alert rules: flexible conditions that can be defined by the user. The conditions can be defined as absolute conditions, e.g., capacity utilization over 100%, or relative conditions, e.g., projected stock >110% max inventory. Alert conditions can be further defined in rules groups with or/and on alert conditions Alert impact key figures: the measure of the impact of the alert: e.g., cost of overstock. Alert charts: the additional analytical charts attached as part of the alert definition to understand the alert situation better. Figure 23 shows the results of alerts generated by running an alert definition on the data. The results show the combinations of resource, location and time for the alert and its impact value. e.g. resource 1011 for location 1010 in week 2020 CW44 has a capacity utilization of 102%. Alerts get notified to the user through emails or are shown in the user interface based on the user’s preference. Users can take further actions from the generated alerts. For example, snoozing the alert or defining a case to manage the mitigation of alert with other stakeholders. Machine Learning for Alert Recommendations Alerts conditions in the alert framework are typically defined as static rules. It works if data is not changing and the thresholds can be well defined. However, many planners find it hard to define the right thresholds and rules for alerts, leading to too little or too many alerts if not defined correctly. Machine learning can perform pattern recognition on the data to detect anomalies and report as alerts. One common outlier technique is to use DBSCAN unsupervised learning clustering method on the data. DBSCAN detects outliers from the data points that do not fit to the clusters. Typically, Planners have to manually provide the lower and upper threshold to detect outliers. However, with machine learning, the system can identify outliers based on the similarity of the key figure value in other time buckets. Fig. 23 User-defined alerts results 42 R. Jandhyala Fig. 24 Insight to action 7 Insight to Action True control tower systems provide insight to action, while allowing supply chain analysts to visualize information, perform analysis, view exceptions and their impacts and finally take actions to resolve the issue (Fig. 24). Scenario: Depleting Inventory For example, the projected inventory in a global inventory position chart shows that in the next quarter, the inventory will be depleting and will go below the safety stock levels. The supply chain analyst needs to analyze what is causing this issue. A drill down into the specific region and time shows that the available supply is less than demand for that quarter. On further analysis, the reason could be that there is a planned factory shut down for maintenance in that quarter and no additional production. Therefore, the demand needs to be fulfilled from the existing inventory. The analysis can further drill down to view which products at which locations are affected and take corrective measures like rebalancing inventory from other locations or using a contract manufacturer. Scenario: Delayed Customer Orders In this scenario, the control tower provides visibility into the sales orders execution data and alerts on the orders that are not confirmed on time or confirmed below the requested quantities. The supply chain planner can then look into the gating factors or root cause analysis related to the orders with the late delivery and then navigate to the source system to change the order dates or quantities to reduce the impact on the customer service levels. An intelligent visibility application brings together all the steps from insight to action in a single user interface, without the planner having to login to different Intelligent Digital Supply Chains 43 Fig. 25 Global visibility with alerts systems containing different sources of information. This drastically improves the productivity and experience of the user as they can see all information in one place and make decisions. The supply chain analyst views the geo-map of the supply chain network with the connections across different distribution centers, plants and customers. There are several alerts displayed on the map in North America for the customer service KPI metric that this person is responsible for. These alerts are caused due to sales order delays, which the analyst would like to further analyze and understand the root cause of the disruptions and its impact on customers. Figures 25, 26, and 27 are adapted from SAP Integrated Business Planning to illustrated the intelligent visibility and insight to action from alerts. Narrowing down on the alerts, the analysts clicks on these 10 alerts and views the individual sales orders and the impacted customers. The analyst drills down into one of the sales orders to get details about the issue and its impact. Here there is a delay of 3 days for a sales order, and the confirmation is at 90% of the requested quantity. The root cause or gating factors for the delay are shown on the same screen (Fig. 26). In this case, there is an insufficient lead time for different components of the finished good. The supply chain analyst can further navigate to the source system based on the document type (Fig. 27), which in this case is the purchase requisition document in SAP S/4 HANA system. In the source system, the order can be changed to re-plan the quantities or dates to meet the customer service levels. In summary, using a single application, the analyst is able to visualize global network view of customer orders and transportations, view alerts for sales order delays, analyze the impacted customers, perform root cause analysis and finally navigate to the source system to take corrective action. This activity, which typically 44 R. Jandhyala Fig. 26 Root cause analysis Fig. 27 Navigate to source system to take action takes days to identify and resolve, is now reduced to hours, increasing the user’s productivity so they can focus on other important activities. 8 Supply Chain Key Performance Indicators Key performance indicators (KPIs) are the measurable primary objectives of business across multiple organizational functions. Supply chains are measured for their performance using KPIs with metrics that are aligned, streamlined and Intelligent Digital Supply Chains 45 transparent across all stakeholders. KPIs can be defined across areas like customer service, demand, supply, inventory, production, logistics and procurement. Supply chain data is vast, with several billions of planning combinations and data points. In order to make a meaningful business representation of the data, KPIs are necessary both to keep track of the progress and measure against the set objectives of the business. The KPIs are reviewed by all roles in the organization from the boardroom to the shop-floor. C-suite executives have well defined KPIs that measure the business growth, e.g., revenue margin and customer service levels. Production planners have KPIs to measure the utilization rate of resource capacity. KPIs should be measurable, transparent, mutually agreed, value oriented and at the right business level to facilitate collaboration, communication and adherence to common goals for an organization. KPIs should be adequate with a proper RACI matrix across the stakeholders, and defining too many or too few KPIs without owners of such KPIs will not add much value. KPIs are visually represented for reporting in dashboards and scorecards. They are usually easy to understand single metrics with visualizations such as speedometer, percentage meter, waterfall chart, traffic lights or single values. They are measured against target values or thresholds. KPI are driven by organizational objectives, as illustrated below: • Sales and marketing: improve promotion effectiveness by 20% by aligning activities across all functional units from commercial, supply chain to finance. • Finance: topline revenue increase by 10% and margin growth by 6% by identifying growth opportunities. • Inventory: reduce out of stocks and balance inventory to reach higher customer service levels, reduce inventory write-offs by 5mi USD by better planning and visibility of products with expiring shelf life. • Demand planning: improve forecast accuracy to 86% from the current 78%, launch new products faster to the market by improving the NPI (New product introduction) process. • Supply chain: reduce the lead time of products from 45 days to 30 days, improve resource utilization from 82% to 95% by streamlining production activities. • Process Improvements: improve planners’ productivity by reducing planning cycle preparation from 8 days to 3 days, increase process automation and improve process efficiency to better utilize time and critical resources. • Customer service: increase customer service levels from the current 92% to a target 96% with improved case fill rates. Common Supply Chain KPIs Some of the KPIs commonly used in supply chain across demand, inventory, supply and production are listed below. These KPIs also bring together the data from planning and execution to measure adherence to the plan. Demand KPIs • Forecast accuracy: it is measured as forecast error (MAPE), where MAPE (mean absolute percentage error) is the average of absolute percentage error between 46 R. Jandhyala forecast and actuals. MAP E = n 100 At − Ft A n t t =1 where A(t) is actual quantity at time t, F(t) is forecasted quantity at time t, and n is the total number of periods • Forecast bias: it measures the average percentage error, which is the deviation of forecasts from actuals. 1 At − Ft n At n MPE = t =1 where A(t) is actual quantity at time t, F(t) is forecasted quantity at time t, and n is the total number of periods Customer Service KPIs • Projected fill rate: ratio of constrained demand (supply) vs customer demand in current and future periods. • Historical case fill rate: measures the ratio of quantities ordered vs. quantities fulfilled across items ordered. • Historical line fill rate: a measure of what percentage of items in the order are fulfilled on time with the right quantity. Inventory KPIs • Days of supply: a measure of the number of days the projected inventory will cover the demand of the subsequent periods Demand Projected inventory Day of supply Day 1 100 300 Day 2 200 200 1.5 0.5 Day 3 300 500 2 Day 4 400 400 2.5 Day 5 100 300 1 • Inventory turns: a measure of inventory cycles for a product which is calculated as the ratio of cost of goods sold to average inventory, both measured in financial valuation • Stock below safety stock: percentage of projected stock below the safety stock for the future periods. • Stock above target stock: percentage of projected stock above the target inventory for the future periods. Intelligent Digital Supply Chains 47 Supply KPIs • Production adherence: a measure of the ratio of planned production (scheduled production) to the actual production from the execution system • Supply shortage: a measure of the missing supply quantity compared to demand forecast across the combinations of products and locations for a given time range. • Capacity utilization: the ratio of capacity usage of the resource to its available capacity. Financial KPIs • Total inventory valuation: total valuated cost of all inventory types including finished goods inventory, raw materials and work-in-progress inventory • Cost of goods sold: cost to serve the goods sold to customers, which includes direct and indirect costs related to materials, labor, production, procurement and transportation. • Inventory turnover: the number of times the inventory should be cycled during a year. It is based on valuated annual inventory and annual COGS i.e. ratio of annual COGS to annual inventory SCOR KPIs SCOR KPIs (APICS SCC 2015) are industry standard KPIs provided by APICS that spans all supply chain business process with a standard set of hierarchical KPIs of around the six primary management processes: plan, source, make, deliver, return and enable. Below is an example of perfect order fulfillment KPI. Perfect order fulfillment: a measure of how the customer orders are serviced based on other detailed KPIs like • Orders delivered in full: a measure of whether an order is delivered to the customer in the right quantities. • On-time performance: a measure of whether orders are delivered within the committed dates and to the correct customer location. • Orders delivered in condition: a measure of whether the orders are delivered with current documentation, e.g., billing, shipping and other documents. KPI Templates Organizations should follow a good template for defining and managing the KPIs. Each KPI should be structured around one focus area, for example, supplier performance with well-defined purpose of the KPI. Below is a guideline for defining KPIs for managing supply chain • Name: choose a name that provides a clear short definition of the KPI without ambiguity. • Definition: provide the business context of the KPI with clarity on the measure (e.g., ratio, percentage, quantity, etc.) and how the results can be interpreted. • Category: tag with the primary focus areas to which this KPI belongs, for example, customer service. • Calculation: provide a precise calculation using mathematical formulas to clearly define the calculation. 48 R. Jandhyala • Roles: define the users who are responsible for the KPI along with other roles players who use this KPI. A RACI matrix of the KPI will help to define the clear accountability of the KPI. • Data source: define the sources of data and transformations required to retrieve and load to the target system where the KPIs are stored or calculated. E.g., sales orders and deliveries with details from order to line items to schedule lines from SAP S/4 HANA, JDA, etc. • Related KPIs: Define how this KPI is related to other KPIs in the same or different categories. For example, KPI of safety stock projection may depend on the KPI forecast accuracy. Interested readers can find more information on supply chain KPI from references (APICS SCC 2015; Ravindran and Warsing 2016). 9 Cognitive Supply Chains Enabled by Technologies in Industry 4.0 Advances in technologies have played a key role and continue to transform supply chains from traditional supply chains to next generation cognitive supply chains. A cognitive supply chain is proactive to sense supply chain disruptions and is capable self-correcting, triggering re-planning and providing intelligent recommendations through advances in artificial intelligence like reinforcement learning where multiple agents work towards the common goal. With technological advancements to store and process the complex and vast amount of growing big data along with flexibility with cloud deployments, supply chains are smarter than ever before to quickly respond in real-time to changing environments. Figure 28 shows the technologies currently available and being used in enterprise business applications to transform the business processes, business models and workspaces. It is adapted from SAP’s positioning of intelligent technologies for the intelligent enterprise. We will look at some of the important technology areas that play a major role in supply chains. • Big data and in-memory computing Advances in big data have made it possible to store a vast amount of supply chain data from multiple sources into data lakes for real time visibility into the supply chain and real time simulations. Big data provides the much needed 4Vs for an enterprise: – Velocity: Frequency of data which can be real time, batch or streaming data. – Volume: store and process petabytes of supply chain data and aggregate results on-the-fly. – Variety: bring together structured and unstructured data. – Value: enable smarter supply chains by deriving value from the data. Intelligent Digital Supply Chains 49 In-Memory computing, Big data Cloud scalability, elasticity, availability Internet of things Machine learning, Artificial intelligence Cloud Platform Blockchain & smart agents Mobile & conversational 3D Printing Autonomous systems & robotics Ubiquitous connectivity, Edge computing Sensors built in/ intelligent things Fig. 28 Technologies driving innovations in supply chains Advances in in-memory computing with columnar databases bring forward the real time enterprise with speed and agility to run complex supply chain operations and do rapid scenario planning. In-memory data management brings together transactional and analytical data in one place without separate systems, redundant data copies or multiple data transformations. Big data and in-memory technologies make it possible for complex analysis of supply chain with data and processes spanning several systems. • Cloud computing Cloud computing, which was once considered a hype, is now a reality and necessity for many enterprises. It gives enterprises the flexibility to onboard new SaaS solutions and extend the existing on-premise landscape with a hybrid approach using the cloud. Several planning and analytical solutions for supply chain management are now available in the cloud, including the ERP systems for storing transactional and master data. Organizations that have adopted cloud see significant benefits with elastic storage and processing, flexible pricing, easy onboarding, and effective maintenance and security provided by the vendors. This allows organizations to focus on important business activities and venture into new business models. Cloud also provides the hyper scalability to store massive amounts of data and provide massive processing power, which helps enterprises run multiple whatif scenarios to determine the best outcome for the business. Further, the cloud vendors for supply chain also provide Platform-as-a-service (PaaS) for complete business suite solutions, including cloud data warehouses, predictive analytics, collaboration, integration, planning, and process orchestration. • Internet of Things Billions of connected, internet-enabled devices now provide a networked community of people, process, data, and things leading new ways to gain insights into a business like never seen before. Internet-of-Things has its place 50 R. Jandhyala and significance in the supply chain, especially in the execution areas like transportation management and digital manufacturing. In the digital manufacturing area, hundreds of IoT sensors are part of the asset, e.g., production resource providing asset intelligence. These sensors monitor the temperature, vibrations, and other performance metrics and can intelligently sense if the asset needs maintenance. In transportation, IoT sensors monitor the state and conditions of goods movement and automate material handling with RFID technology across several logistic hand-over points without human intervention. For example, in medical devices transportation, IoT sensors can monitor the temperature fluctuations and severe vibrations that can damage the equipment. IoT is an essential technology in the Industry 4.0 revolution of supply chains. • Blockchain Blockchain provides the transparency, traceability, and auditability of all documents and events that happen in the supply chain for material movements. There have been good use cases for blockchain in supply chains in the areas of agriculture, logistics and life sciences. For example, in agriculture, blockchain helps track every stage of the process for the fresh food items to reach consumer, and in an event of outbreak, trace back to the origin. In logistics, blockchain helps track every movement of material across logistic providers, shipping ports and inter-continental trading hubs to instantly track and respond to delays. • Machine learning, artificial intelligence and robotic process automation Intelligence and automation elevate supply chains to focus on high value activities. The digital transformation of supply chains, together with advances in big data, algorithms and hardware, have enabled the new AI era of the cognitive supply chain where systems learn from the data to augment human decision with intelligent recommendations. Supply chain systems have traditionally relied on statistical and deterministic models for forecasting, supply network optimization and several other planning and execution processes. With machine learning and AI techniques for supervised learning, un-supervised learning, reinforcement learning and deep learning, systems can analyze patterns in the data, correlate with external data and events and augment or replace the traditional planning with cognitive planning. For example, demand sensing augments the statistical demand planning process by sensing short term demand based on pattern recognition of sales orders, sales history, demand history and external data like weather and point of sales. Using robotic process automations, supply chain users can deploy bots that can carry out mundane tasks of users, thereby freeing up their time for other productive activities. The bots can carry out tasks like reading email for new procurement requests, logging into multiple systems for data entry, solving supply chain alerts, and trigger workflows. A cognitive supply chain self-learns from the patterns in the data and provides predictive and proactive insights to the users to make timely decisions before disruptions occur. It provides: Intelligent Digital Supply Chains 51 1. Automation to eliminate most of the manual processes and interventions based on exceptions 2. Alerts and suggested actions 3. Automatic execution of corrective measures, self-learned from the users’ actions. One of the applications of machine learning and automation in cognitive supply chains is in the area of analytics, transforming the analytics from Descriptive analytics to |______ Predictive analytics to |________ Prescriptive analytics Descriptive Analytics Many organizations see the need to view and report on the data and have invested over years to build analytics using Data Warehouse platforms and powerful analytical and visualization tools. These are descriptive analytics with fixed representation and visualization of data. Figure 29 shows a dashboard which has both structured and unstructured data, as adapted from SAP for illustration. By integrating customer sentiment data, the User is able to see how the negative perception of the product is affecting the customer satisfaction. Such analytics help users to instantly see the related data together to make some inferences. Predictive Analytics Predictive analytics provide smart data discovery tools for predictive insights into the data and predicting future outcomes using Machine Learning. This helps with for proactive analysis rather than being reactive. For example Figure 30 is based on the customer sentiments analysis, it is predicted what the future customer satisfaction will be and how it affects the revenue and profitability. This figure is adapted Fig. 29 Descriptive Analytics 52 R. Jandhyala Fig. 30 Predictive Analytics from SAP Integrated Business Planning to illustrate predictive analytics. Planners can then run different strategies to restore back the customer satisfaction while maintaining profitability. Prescriptive Analytics A cognitive supply chain provides the next level of intelligence by sensing the impact of the signals and applying intelligence to come up with alerts along with prescriptions or recommendations to solve the alerts. The system evaluates multiple what-if scenarios and business impact analysis to recommend the best path forward using techniques like simulation, optimization and machine learning. In Fig. 31, the system recommends the course of action for supply planning by recommending make-and-buy vs buy options along with impact analysis on supply chain costs and a ranking for the recommendation. This figure is adapted from SAP positioning of recommended actions to solve supply chain alerts. Evaluation of Supply Chain Assumptions, Risks and Opportunities Supply chains have vast amounts of data distributed across multiple functional systems. To manage the business outcomes in a supply chain, a well-supported decision-making process driven by managing both qualitative and quantitative information is required. With well-managed process to capture assumptions, risk and identify new opportunities, organizations can quickly analyze their data and run meaningful simulations of the business outcomes and reduce vulnerabilities. Intelligent Digital Supply Chains 53 Fig. 31 Prescriptive Analytics Examples of qualitative business drivers include events, signals, assumptions, risks, opportunities and promotions. Signal New qualitative information, including measurable ones like new sales orders, and non-measurable ones like a bad product review or market trend. These signals, when qualified as impactful, are considered in a planning cycle. Assumptions Qualitative business drivers on expected business conditions and outcomes. These are defined and consolidated across all functional units and tracked for changes or deltas in each planning cycle. Example: market share steady for Q1 in EMEA region at 18%. Risks and Opportunities Risks capture supply chain vulnerabilities, e.g., supplier risk, and opportunities are new chances, e.g., an increase in market share with competitors going out of business. These drivers are clearly defined for each planning cycle with its impact, duration, planning level and qualitative attributes like status, inclusion-to-plan, probabilities. Compared to assumptions, the risks and opportunities are well defined with business impact, duration and probability. These are reviewed collaboratively in business review meetings and the agreed risks and opportunities are included into the plan. Events Events are both internal and external activities that are known upfront (e.g., sports events and holiday seasons—Christmas) or unknown (e.g., supplier warehouse affected by a hurricane). Events can also be cyclic, which repeats each 54 R. Jandhyala year. Promotions can be thought of as a variant of an event with a short time span, e.g., 2 with 1 offer and back-to-school promotions. A well-orchestrated process for managing business drivers has the following benefits: • Captures all qualitative drivers like internal or external and short term or long term across all functional units • Reacts to vulnerabilities or new opportunities much quicker. • Increases transparency and trust across different functional units and stakeholders. • Consolidates drivers across top-down drivers from strategy and finance, with operational drivers entered by local and regional planners and cross-functional across different hierarchies. • Increases productivity by capturing all influencing drivers that impact the supply chain in one place. • Collaborates on decisions making on which drivers to be included in planning cycles • Improves business review meetings by evaluating multiple scenarios of which drivers to include in the plan, which is otherwise done offline with a toll on time and resources. Process for Managing Business Drivers A sample business process of managing business drivers, including assumptions, risks and opportunities is shown in Fig. 32. This can be different for each organization based on its business purpose. Assumptions are typically defined once a year with collaboration across all the functional units, including commercial, supply chain, finance and strategy. Assumptions are defined at an aggregate hierarchical level, example product category along with the influencing drivers, example growth, market share, competition etc. for each month or each quarter. During the monthly review meeting, the assumptions are reviewed and updated with any changes based on the market conditions and actual performance of the business. Risks and opportunities are defined for each planning phase i.e., demand, supply, financial reconciliation, etc. Next, the quantitative plans are adjusted to include the agreed risks and opportunities. Further, any changes to the forecasts are tagged with the change in assumption. This provides a context for the quantitative changes by associating them with the assumption changes. During the review meetings, an assumption changes report is generated, which brings together the qualitative influencing factor and the quantitative change, e.g., forecast change associated with it. SAP integrated business planning product provides the required tools and functionality to easily manage the business drivers. Figure 33 is an example of managing risks and opportunities in a driver-based planning application of SAP IBP that provides the required user interface, sample data model and best practice content to easily manage risks and opportunities and collaboratively include them in the plan. Intelligent Digital Supply Chains 55 Fig. 32 Business process to manage assumptions, risks and opportunities Fig. 33 Planning views to manage different business drivers The planning views provide the flexibility to define a business driver e.g., risks and opportunities across multiple planning hierarchy levels, time ranges and driver impact key figures for the functional area. For example, there can be demand risks and opportunities defined for the North America region for current and future quarters; supply assumptions defined for all regions for the next 2 years. From the planning view, which defines the structure of the business drivers, planners can navigate to view and define the individual drivers. For example, below is a summary view of all risks and opportunities defined at the hierarchical level product family and customer region. Each driver has the following properties: 1. Qualitative information about the driver: this includes name, description, planning cycle and type. 2. Hierarchical level: a driver can be defined for one or more hierarchical levels, e.g., risk 1 is for product family 1 and product family 2 for all customer regions, and risk 2 is for product family 3 for EMEA region. 56 R. Jandhyala 3. Time range: each driver can be defined for different time ranges at a defined time granularity, e.g., Jan–May 2020. 4. Impact key figures: these are the qualitative measures representing the impact of the key figure. For example, opportunity 1 has an impact of 5mi USD in revenue. 5. Qualifying attributes: these include driver attributes like budgeted, status, probability and include-in-plan. The drivers can be defined in one go at multiple planning hierarchical levels and time ranges, with a qualitative value for the driver impact key figure. The key figure value can also be adjusted for different time periods or planning combinations. For example, opp1 has higher weightage of revenue in the later quarter compared to the first 3 quarters. The summary view lists all the risks and opportunities along with analytical charts to view the impact of the drivers on planning data like top 5 risks, top 5 opportunities, alerts on risks and opportunities above thresholds (Fig. 34). Planners can review all the risk and opportunities in one go and must maintain drivers by changing attributes like probability and status or carry forward drivers to the next planning cycle. Qualified drivers with the right granularity of planning attributes, which are budgeted and financially viable, and with higher probability, can be included in the plan by changing the attribute value include-in-plan to 1, and its effects can immediately be seen in the planning data. The contribution of the risk or opportunity can immediately be seen in the supply plan. For example, risk1, risk 2 and opp1 are included in the plan. These drivers are defined at an aggregated level of the product family and customer region than the final forecast key figure, which is at the product/customer level. The contribution of these three drivers is disaggregated from aggregate level to product customer level Fig. 34 Risks and opportunities summary view Intelligent Digital Supply Chains 57 Fig. 35 Contribution of risks and opportunities on supply plans of the final forecast key figures and added or negated from the planned forecast to arrive at the final forecast value that represents the plan considering risks and opportunities, as illustrated in Fig. 35. 10 Resilient Supply Chains Resilient supply chains require intelligent visibility and digitization of supply chains. A holistic approach is required than a piecemeal approach to gain visibility into the end to end supply chain connecting functional domains like planning, manufacturing, logistics and execution. This is required to determine conditions that need to be analyzed, simulate the impact and solutions for those conditions, apply intelligence to evaluate alternatives that are financially viable, and finally, deploy the solution without impacting the current execution. Supply chains are more vulnerable than ever before, and businesses face unforeseen macro and micro supply chain disruptions. Examples include new tariffs, trade wars, Brexit, hurricanes, pandemics like COVID-19, SARS, etc. Further planned events such as seasonal sales, sports—FIFA, Superbowl, etc. and other internal and external events with shifting market demands and supply have a big effect on the supply chains. Other examples include: • • • • Impact on late shipment on customer order fulfillment Impact of unforeseen asset downtimes on production Impact of EPA regulations on logistics and distribution of goods Impact of raw material shortages e.g. impact of fires in Canada influence the paper pulp used for toilet papers. • Impact of pandemic like COVID-19 on medical supplies and essential services. Such planned and unplanned events affect availability of supplies and lead to long revival times, increased financial risk and labor shortages, for example, COVID social distancing impacting production, as distancing rules need to be followed at production lines. Figure 36 shows examples of COVID-19 impact and other supply chain events that make the supply chains vulnerable. These figures are adapted from SAP’s positioning of Resilient Supply Chain. 58 R. Jandhyala Fig. 36 Supply chain vulnerabilities and impact Resilient supply chains proactively determine supply chain disruptions, simulate events and find alternatives, and recover from disruptions. Goals of intelligent digital systems in resilient supply chains include: • Early warning systems that show vulnerabilities in supply chain. • Faster what-if scenarios and simulations to find the best alternative. • Use of the abundant supply chain knowledge from historical data combined with planner’s actions and enterprise goals. • Alignment of operations and finance to find the best financially viable solution. • Application of intelligence to supply chains by co-relating data from multiple sources to identify supply chain relationships, risks and opportunities. • Mitigation of disruptions in a timely manner with transparency and collaboration across stakeholders. • Signal based management of short-term and long-term impacts. • Manual to automated search and discovery of supply chain vulnerabilities. • Continuously test the robustness of the supply chain by simulating failures and tracking how supply chain respond to such disruptions. Intelligent Digital Supply Chains 59 • Risk sensing to reduce risk from potential disruptions. • Resilient supply chains require intelligent visibility and digitization of supply chains that was discussed in the previous sections. A holistic approach is required than piecemeal approach to gain visibility into the end to end supply chain connecting functional domains like planning, manufacturing, logistics and execution. This is required to determine conditions that need to be analyzed, simulate the impact and solutions for those conditions, apply intelligence to evaluate alternatives that are financially viable and deploy the solution without impacting the current execution. Examples of resilient supply chains across different Industries Healthcare: • • • • Manage allocation and inventory of available PPE equipment’s inventory Planning of critical medical supplies during pandemics and natural calamities. Telecommunications: Faster response during catastrophic events or pandemics e.g. during COVID-19 where there was greater demand for network communications • Manage available inventory for critical customers. For example, higher bandwidth allocation for government and medical workers. Agriculture: • Increase shelf life of products from farm to fridge. • Manage production with plant shutdowns and fluctuating consumer demand. e.g. Sugar demand surged during COVID pandemic where people during lockdown were consuming more sugar. • Better sustainable products with natural, organic ingredients. Consumer products: • Alternate sourcing for raw materials. • Alternate modes of transport (rail, road, sea or air) to supply materials with high demand. e.g. toilet papers, hand sanitizers, etc. • Strategies for long term vendor determination. • Faster product interchangeability to offer competitive products and optimally consume supply from similar products. Food and Beverages: • Visibility of inventory in global supply chain and balance available inventory to reduce obsolescence. For example, scarcity of water bottles during hurricane or increased demand for chips and soda during gaming season. • Scenario planning to determine alternate sources for packaging if packaging factory is hit by calamity like fire. 60 R. Jandhyala Oil and Gas • Rapid scenario planning with fluctuating crude oil prices • Demand collaboration across extended supply chain to determine projects for drilling, fracture and oilfield rigs. • Optimal productions and transportation of oil by-products. For example, during pandemic, oil reservoirs were placed on containers in sea. High-Tech • • • • Real time visibility of critical suppliers for materials. Evaluate alternate lead times and production lot sizes Distribution planning to manage service levels of critical supplies. Financial impact of raw material price changes 11 Collaborative Enterprise Planning: Integrated Supply Chain and Financial Planning An integrated operations and financial planning system helps enterprises instantly gain visibility into the supply chain plans with its financial impact and vice versa. However, the financial planning systems and supply chain systems are separate and disconnected from each other, each with a different purpose and ownership. The finance data focuses more on valuation (dollarized amount e.g., revenue, costs and margin) with data structure based on financial account hierarchies at an aggregated company, product line levels. On the other hand, supply chain data focuses on volumes or quantities (e.g., demand forecasts and production quantity), and the supply chain is a network model with material flow across connected locations and bill of materials. A Collaborative enterprise planning solution supports an integrated financial, supply chain and commercial process. It consolidates budgets and financial plans with commercial and operations plans, balancing supply, demand and inventory with financial goals plans across the organization. Use Cases for an integrated finance and operations system include the following: – – – – – – – Strategic planning Investment plan on new production line Headcount decisions to increase shifts for production Capex planning on retiring an aging asset Profitability impact on price changes for raw materials Impact of mergers and acquisitions or sell-off Trade effectiveness and tax changes The business benefit of integrating the two systems is to provide an output of a single approved plan (financial/operating) as a blueprint for all stakeholders Intelligent Digital Supply Chains 61 Fig. 37 Planning types and horizons for finance and supply planning Fig. 38 Supply Chain Costs-to-Serve (internal and external) to follow, which positively impacts the balance sheet and profit and loss statements of the business. Most companies follow these planning cycles shown below for strategic planning, which is 3–5 year long term plan followed by an annual operating plan or budget plan followed by sales and operations planning for the Rolling 18 months. An example of the planning processes is shown in Fig. 37. Financial planning systems need the volume-based supply chain plans for running processes like budget planning and product cost planning, and the supply chain planning systems need the valuation based financial plans for consensus demand planning and financial reconciliation. Supply chains have financial costs associated with each step to fulfill customer demand, as shown in Fig. 38. These include raw material costs, production costs, productive costs, direct and indirect costs, transportation costs and storage costs. Supply chains employ profit optimization techniques to arrive at a profitable costeffective strategy to fulfill customer demands. Figure 39 shows a consensus planning view in SAP Integrated Business Planning consolidating inputs from all functional units -sales, marketing, demand planning, finance to arrive at a consensus demand plan. AOP represents the annual operating 62 R. Jandhyala Fig. 39 Financials in supply chain plans plan. Both volumes and values of different plans are recorded or calculated, and the gap between AOP and consensus plans is visible month to month. Planners have a consolidated view of the complete demand picture and can make adjustments to close the gap to the financial targets. Revenue and profit are calculated automatically based on adjustments to the volumes. The scope of financials in supply chains is to provide directional finance and valuation for supply planning decisions. These include: • Financial projections (valuation of the plans) that convert volumes to values and report at different aggregation levels • Currency conversions including simulations on price and exchange rates changes • Rolling revenue and gross margin projections • Planned price and costs as input at different aggregate levels than supply chain plans. • Budget plans or annual operating plans as financial targets to drive operation plans • Calculations of price based on revenues and volumes • Weighted average of price and costs calculations • Basic cost rollups for cost-to-serve calculations The overall process for finance and supply chain integration is shown in Fig. 40 across tactical annual budget planning and operational rolling monthly revisions of the plan. This figure is adapted from SAP’s positioning of collaborative enterprise planning between finance and supply chain solutions. Annual Financial Planning 1. Tactical annual financial planning is performed in financial planning systems for the next fiscal year combining inputs from strategic planning, prior year plans and operations plans from a supply chain that consolidates demand and supply plans for the next 1–3 years. Intelligent Digital Supply Chains 63 Fig. 40 Integrated finance and operational process 2. Supply chain systems send the operational plan to financial planning systems as a starting point for granular financial planning (e.g., detailed annual budget—sales units by region, cost assumptions and investments in capacity) that integrates the operational plan. 3. The resultant annual operating plan is sent to supply chain systems as a financial target for consensus and constrained demand planning to evaluate scenarios that bridge the gap to meet financial targets. Operational Rolling Financial Planning 1. The annual operating plan is sent from financial planning to supply chain planning systems once a year. Iterations occur monthly in the supply chain until a consensus is reached on a single plan considering the financial budget and constrained supply plans that allocate the critical resources, inventory and materials most efficiently to satisfy customer demands while maintaining high profit margins and working capital. 2. The process in supply chain starts with arriving at a consensus demand plan in S&OP based on sales, marketing, demand and financial plans followed by constraining the plan based on profit optimization considering supply chain constraints and costs, along with business assumptions, risks and opportunities (Kusters et al. 2018). 64 R. Jandhyala 3. Every stage of the process in the supply chain also includes financial projections like revenue, margin, costs for the demand and supply plans to see both the volume and value for the plans. 4. Finance and operations collaborate monthly or quarterly on variance analysis reporting (plan vs. actuals) to reforecast and model changes in the business, industry or economy. 5. The revised demand, supply and inventory quantities are sent to the financial planning system for monthly/quarterly revision of the plan. This includes running processes like sales revenue planning, cost and activity planning, product cost planning, profitability planning, and P&L planning. The net result is an aligned profit and loss, balance sheet, and cash flow statements with the supply chain plan. 6. The results are revised financial targets that are sent to supply chain for visibility into the latest P&L. The revised financial plans, i.e., Q2 revision, Q3 revision, Q4 revision are the new financial targets for the supply chain. References APICS SCC. (2015). The supply chain operations reference model (SCOR® ). Chopra, S., Meindl, P., & Kalra, D. V. (2013). Supply chain management: Strategy, planning, and operation (Vol. 232). Boston, MA: Pearson. Kepczynski, R., Ghita, A., Jandhyala, R., Sankaran, G., & Boyle, A. (2019a). Enable IBP with SAP integrated business planning. In Implementing integrated business planning (pp. 23–110). Cham: Springer. Kepczynski, R., Dimofte, A., Jandhyala, R., Sankaran, G., & Boyle, A. (2019b). Implementing integrated business planning (Management for Professionals). Cham: Springer. Kusters, J., Jandhyala, R., Mane, P., & Sinha, A. (2018). Sales and Operations Planning (S&OP) with SAP IBP (SAP PRESS). Ravindran, A. R., & Warsing, D., Jr. (2016). Supply chain engineering: Models and applications. Boca Raton, FL: CRC Press. Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design Aswin Dhamodharan and A. Ravi Ravindran 1 Introduction Closed Loop Supply Chain (CLSC) management is defined as ‘the design, control and operation of a system to maximize value creation over the entire life cycle of a product with dynamic recovery of value from different types and volumes of returns over time’ (Guide and Li 2010). According to the authors, a business process perspective on CLSC consists of three subprocesses, namely: (i) Front-end that deals with product return management, ensuring the availability of products at required quality and quantity. (ii) Engine that deals with operational issues in remanufacturing, ensuring profit from remanufacturing. (iii) Back-end that deals with market development of remanufactured products, ensuring demand for recovered products. OEM has to make two major strategic decisions at the Front-end of CLSC, namely, (i) identify the most profitable channel of product recovery and (ii) identify the profit maximizing remanufacturing market that OEM should participate in. We explicitly model the optimal collection of remanufactured products through suppliers in a dynamic manner across multiple time periods over the product life, accounting for demand, quality and remanufacturability of returned products. A. Dhamodharan () Tesla Motors, San Carlos, CA, USA e-mail: adhamodharan@tesla.com A. Ravi Ravindran Pennsylvania State University, State College, PA, USA e-mail: axr32@psu.edu © Springer Nature Switzerland AG 2021 S. Srinivas et al. (eds.), Supply Chain Management in Manufacturing and Service Systems, International Series in Operations Research & Management Science 304, https://doi.org/10.1007/978-3-030-69265-0_3 65 66 A. Dhamodharan and A. Ravi Ravindran At the back-end of the CLSC, the OEM has to (i) make pricing decisions for new and remanufactured products based on the consumer willingness to pay (WTP) and (ii) estimate the demand for new and remanufactured products throughout the product lifecycle. We use the extended Bass model (Debo et al. 2006) to model total demand for the product, and cannibalization of new product demand is reflected in pricing decisions of new and remanufactured products. We use pricing decisions for new and remanufactured products based on Abbey et al. (2015a, b). We develop a multi-period CLSC network design planning model, incorporating product life cycle for a supply chain consisting of three stages: (i) OEM manufacturing facility, (ii) distribution centers, hybrid facilities and recovery centers and (iii) retailers. At the strategic level, the OEM has fixed locations for manufacturing plants and retailers, and potential locations for distribution centers, hybrid facilities and recovery centers. In the forward flow, OEM manufactures new and remanufactured products and distributes them through warehouses and hybrid facilities to satisfy demand at the retailers in fixed locations. In the reverse flow, retailers collect commercial returns, end of use returns and end of life returns from consumers for a collection cost, provided the return has a certain level of functionality. These returns are then sent to hybrid facilities and recovery centers, where they are disassembled, inspected and stored. The OEM’s manufacturing facilities pull these returns from the hybrid facilities based on the demand for remanufactured products. At the tactical level, we consider an uncapacitated model, where demand for new and remanufactured products follow an extended Bass model (Debo et al. 2006). We consider a segmented market with quality-sensitive and price-sensitive consumers based on the empirical evidence from Abbey et al. (2015a). We assume that the WTP functions for both quality-sensitive and price-sensitive consumers overlap. We use monopolist pricing decisions for new products in all periods (Abbey et al. 2015b). When making a pricing decision on remanufactured products, we use monopolist pricing expression (Abbey et al. 2015b) only when there are enough available returns to satisfy the demand. Under constrained availability of returns, we use a myopic pricing decision, where the OEM maximizes profit from selling remanufactured products for that period only. We also penalize any lost demand due to shortage of returns. In the literature review, we briefly discuss contributions to dynamics of product lifecycle; consumer perception of new and remanufactured products, the demand cannibalization for new products from remanufactured product sales and finally, pricing decisions under remanufacturing. We introduce the product life cycle optimization model and use a case-study to show that the OEM’s approach to supply chain network design can significantly impact the profitability in the reverse supply chain over the product lifecycle. Finally, we present the conclusions from sensitivity analysis on various parameters in the model. Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 67 2 Literature Review 2.1 Product Life Cycle Dynamics We use diffusion based models to study product lifecycle dynamics. Bass (2004) has developed a growth model to forecast demand for consumer products, called the Bass model. The Bass model is based on the behavioral assumption that the probability of a new purchase of a product at any time depends on the number of previous purchases of the product. The model is empirically tested and found to accurately predict the highest number of sales in a period and its timing, for various consumer durables, predominantly between 1950 and 1960. Debo et al. (2006) are the first to consider product life cycle dynamics in a closed-loop setting. The authors extend the Bass diffusion model to: (i) study life cycle demand of new and remanufactured products and (ii) analyze investments in flexible capacity. The authors find that slow diffusing products are best suited for remanufacturing, while fast diffusing products and products with repeat sales, make investment in flexible capacity valuable. They also comment that the right time for the OEM to introduce remanufactured products is after the sales for the new product has peaked. In addition to the closed-loop setting and the factors studied in this work, our model additionally considers the evolution of product pricing over the product lifecycle, since profit is a function of pricing and demand. There are very few studies in the literature that explicitly consider supply chain decisions, including pricing, across the life cycle of a product, in a closed loop setting. To the best of our knowledge, Chen and Chang (2013) is the only study that analyzes pricing decisions across the life cycle of a product, but does not model the product demand. Papers that analyze specific CLSC problems, such as, Savaskan et al. (2004), Ferrer and Swaminathan (2010), Toktay and Wei (2011), consider one or two-period product life cycle to keep their models tractable. On the other hand, papers analyzing broader CLSC problems, such as, Jayaraman (2006), Min et al. (2006), Pazhani and Ravindran (2014), do not model demand from the product life cycle view. 2.2 Consumer Perception of Remanufactured Products Ovchinnikov (2011) shows empirical evidence that, because of the lower price of the remanufactured product, a new market segment is ready to purchase remanufactured products. Ovchinnikov (2011) also provides empirical evidence that, as the price of remanufactured product increases, shifting of consumers from new product purchase to remanufactured product purchase follows an inverted U shape. Pang et al. (2015) analyze data on purchases of electronic products available in eBay, UK. The authors determine seller identity and reputation, length of warranty period, supply of remanufactured products, proxy demand, duration of sale and end day 68 A. Dhamodharan and A. Ravi Ravindran of sale as significant factors that determine the price differential between new and remanufactured products. While the studies above characterize customer WTP, they lack extensive empirical evidence. In our work, we use the market structure proposed by Abbey et al. (2015a, b), who conduct multiple studies to understand consumer perceptions of remanufactured products, including technological, household and personal products. The authors find that product quality is significantly more important than the discount offered in all the studies. They also find that brand equity has insignificant impact and educating consumers on remanufacturing process does not help in alleviating their negative perceptions. Based on their extensive experimental results, the authors conclude that consumers fall into one of the two categories: (i) A new product only segment, where the consumer always has a significantly higher preference for a new product. Even under extreme discounts for remanufactured products, the consumer prefers only new products. For all future uses, we refer to this consumer segment as Quality-sensitive consumers. (ii) An indifferent segment, where the consumers are indifferent between new and remanufactured products. When both new and remanufactured products are presented at the same discount level, the consumer segment only shows a slight preference for new products. For all future uses, we refer to this consumer segment as price-sensitive consumers. We follow the consumer behavior model of Abbey et al. (2015b), overlaying it on the demand model, and explicitly model the OEM pricing decisions over the product lifecycle. 2.3 Demand Cannibalization Introduction of remanufactured products in the marketplace affects the demand for new products, called demand cannibalization. The majority of models in the CLSC literature assume no cannibalization, which is not always true. Guide and Li (2010) provide the first empirical evidence on the effect of product cannibalization. Ovchinnikov (2011) argues that remanufacturing decisions based on consumers’ WTP for remanufactured product could be misleading, as lower price of remanufactured products does not always result in increased sales. This is the first study to quantify product cannibalization, based on a behavioral study for cell phones, and employ the observed cannibalization function into the model. However, the focus of this study does not include network design and demand considerations over the product lifecycle. Since we observe significant evidence for the cannibalization of new product demand in the literature, we consider the cannibalization effect through the OEM pricing strategy from Abbey et al. (2015b). Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 69 2.4 Pricing Under Remanufacturing Majumder and Groenevelt (2001) initiated a stream of papers on the pricing strategy for remanufacturing. The majority of the articles begin with the assumption of a theoretical model for consumer behavior (Abbey et al. 2015b). Hammond and Beullens (2007) are the first to extend the pricing model to CLSC. The authors consider oligopolistic CLSC under the Waste Electrical and Electronic Equipment (WEEE) directive. They derive the Nash equilibrium for the supply chain and find that the legislation encourages reverse supply chain activities. Abbey et al. (2015b) compute the optimal pricing strategy based on an empirically observed market structure. The authors initially derive optimal pricing expressions of a two-period model, where the OEM sells only new products in the first period and sells both new and remanufactured products in the second period. The main assumption in the initial model is no constraint on the availability of product returns. The authors then impose this constraint and propose an algorithm to compute the optimal pricing strategy. We use both the market structure and the optimal pricing strategy given by Abbey et al. (2015b). We observe that OEM’s pricing strategy influences the sales of new and remanufactured products. Also, the quantity of new products sold impacts the quantity of returns. Hence, we analyze the OEM’s pricing strategy at the network design level. 3 Methodology 3.1 Product Life Cycle Optimization Model for CLSC We consider a single product, multi-period CLSC network model for an OEM, as shown in Fig. 1. The location of manufacturing plants (MP) and consumers (C) are fixed and known. In the forward supply chain of the network, MPs produce new and remanufactured products that flow through distribution centers (DC) and hybrid facilities (HF) and are sold to consumers (C) through retailers (R). In the reverse supply chain, retailers (R) collect consumer returns and distribute them to the recovery centers (RC) and the hybrid facilities (HF), where they are disassembled, tested, sorted and then sent to the manufacturing plants (MP), for repairs and remanufacturing. The network optimization problem determines the optimal locations for the DCs, HFs and RCs that maximize the profit of the CLSC network, given fixed locations for all MPs and retailers. We consider inventory holding cost, transportation cost, fixed cost of opening the facilities, and shortage cost. We also consider the revenue from selling new and remanufactured products and the disposal cost. The following subsection describes the OEM’s pricing decision that affects product demand. Initially, the OEM sells only the new product in the market. After a few periods, the OEM introduces the remanufactured product in the market. The optimization model generates demands based on the demand 70 A. Dhamodharan and A. Ravi Ravindran Distribution Centers (DC) Manufacuring Plants (MP) Hybrid Facilities (HF) Retailers (R) Consumers (C) Recovery Centers (RC) Fig. 1 Structure of CLSC network examined in PLCOM. Solid arrows represent the forward flow of new and remanufactured products from the manufacturer to consumers. Dotted arrows represent the backward flow of returns from consumers to the manufacturer model (Appendix 2) that reflects product life cycle parameters. The optimization model decides the optimal demand fulfillment and the optimal selling price based on the pricing model (Appendix 1). The output of the optimization model gives the optimal network design and the optimal distribution plan that maximizes the profit over the product life cycle. 3.2 Product Life Cycle Optimization Model (PLCOM) for CLSC In this section, we present an integrated optimization model called product life cycle optimization model (PLCOM), for designing an optimal CLSC network. The PLCOM consists of three models. The functions of these models and their interactions are shown in Fig. 2. The major features of the PLCOM for CLSC are as follows: 1. A framework that applies the pricing model (Appendix 1) to decide the selling price of new and remanufactured products. These pricing decisions are exogenous to the optimization model. 2. Integration of pricing decisions with the demand model (Appendix 2) that accounts for the product life cycle. Thus, we compute the demand and the selling price for new products and remanufactured products, exogenous to the optimization model. 3. Integration of the demands and prices computed in the previous steps with a network design optimization model. Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 71 Fig. 2 Product life cycle optimization model for CLSC 3.3 Assumptions of the Integrated Optimization Model The following are the assumptions of the integrated optimization model: 1. All types of returns, such as commercial returns and end of use returns, are modeled as consumer returns. Remanufacturability of all consumer returns is assumed to be constant over the entire planning horizon. 2. Lead time for both manufacturing and remanufacturing of the products is zero. This assumption can be easily relaxed to study the effect of lead time. 3. Transportation time across the supply chain is zero. 4. Demands at retailers are generated using both the pricing model and the extended Bass model, for each period. 5. There are no capacity restrictions at the manufacturing plants for new and remanufactured products. 6. Demand for new products must always be fulfilled. 7. Shortages are allowed for remanufactured products. However, all shortages are treated as “lost-sales”. 8. The fraction of price-sensitive consumers and fraction of quality-sensitive consumers remain constant throughout the planning horizon. 9. Consumer WTP functions (α 1 + β 1 θ 1 ) and (α 2 + β 2 θ 2 ) are constant over the entire planning horizon. 10. Pricing model assumptions (Section Appendix 1) are (i) cn > cr , (ii) α 1 > α 2 and (iii) α 1 + β 1 > α 2 + β 2 72 A. Dhamodharan and A. Ravi Ravindran 11. Consumer returns for remanufactured products are not allowed. 12. Proportion of consumer returns that is remanufacturable is constant. 13. Only a fixed proportion of consumers make a repeat purchase. The rest of the consumers are assumed to not buy the product again during the product lifecycle. 3.4 Optimization Model Sets: N: planning horizon of the optimization model, t = 1,2,..,N MP: set of all manufacturing plants in the network DC: set of all potential distribution centers HF: set of all potential hybrid facilities RC: set of all potential recovery centers R: set of all retailers in the network Parameters used from integrated pricing and demand model (Appendix 3): market size: the total size of the market considered. marketsize is normalized to 1 innov: innovation coefficient of the product that creates new demand. The coefficient is expressed as a fraction of the total market size imitation: imitation coefficient that creates new demand through word-of-mouth effect from existing customers. This coefficient is also expressed as a fraction of the total market size marketexpt : market size expanded in period t, expressed as a fraction of total market size markett : market size at the end of period t, expressed as a fraction of total market size ϕ: the proportion of returns that can be remanufactured (quality of consumer returns) residencej : the fraction of current product owners, markett , whose new product will fail at period j repeatt : demand created from repeat purchase by current product owners in period t. This unit is a fraction of markett L: product lifetime newprdownt : number of consumers who own the new product in period t returnst : number of consumer returns of new product in period t. Note that we do not consider returns of remanufactured products. The consumer return may be a commercial return or end of life return Qnewt : optimal production quantity for the new product in period t Qremt : optimal production quantity for the remanufactured product in period t Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 73 λ: the fraction of consumers that only buy a new product, referred to as Qualitysensitive consumers 1-λ: the fraction of consumers that only buy a cheaper product, referred to as Pricesensitive consumers α 1 + β 1 θ 1 : willingness to pay function for quality-sensitive consumers α 1 : intercept of willingness to pay function of quality-sensitive consumer β 1 : the slope of willingness to pay function of quality-sensitive consumer θ 1 : variable between 0 and 1 to give cumulative demand for new product α 2 + β 2 θ 2 : willingness to pay function for price-sensitive consumers α 2 : intercept of willingness to pay function of price-sensitive consumer segment β 2 : slope of willingness to pay function of price-sensitive consumer segment θ 2 : variable that varies between 0 and 1 to give cumulative demand for new product innov: innovation coefficient of the product imitation: imitation coefficient of the product ϕ: the fraction of consumers that repeat purchase of a new product remtime: time to remanufacture a product return repcust: the fraction of current product owners making a repeat purchase (treated as a constant) cn : manufacturing cost for new product intro: time period at which remanufactured product is introduced Cost components: cn : manufacturing cost for new product pn : selling price for new product pr : selling price for remanufactured product cn : manufacturing cost of a new product cr : remanufacturing cost dispcost: unit cost of disposing a product hc : unit cost of holding a returned product in inventory per period hr : unit cost of holding a remanufactured product in inventory per period sc: cost of not fulfilling a demand per unit fw : fixed cost of opening a DC fh : fixed cost of opening a HF fr : fixed cost of opening a RC tf 1mp,dc,s: unit transportation cost from a manufacturing plant mp, to retailer s, through a distribution center dc tf 2mp,hf,s: unit transportation cost from a manufacturing plant mp, to retailer s, through a hybrid facility hf tb1s,rc,mp: unit transportation cost from a retailer s, to a manufacturing plant mp, through a recovery center rc tb2s,hf,mp: unit transportation cost from a retailer s, to a manufacturing plant mp, through a hybrid facility hf 74 A. Dhamodharan and A. Ravi Ravindran Capacity components: cap1mp: maximum capacity of manufacturing plant mp cap2dc: maximum capacity of distribution center dc cap3hf : maximum capacity of hybrid facility hf cap4rc: maximum capacity of recovery center rc Demand at retailers: ωs : fraction of the total demand at retailers γ s : fraction of consumer returns at retailer s Qnews,t : demand for the new product at retailer s during period t, and is given by: Qnews,t = (Qnewt ) ωs ; where ∗ For t < intro, Qnewt = qn1 marketdemandt ∗ For t ≥ intro, Qnewt = qn2 marketdemandt Qrems,t : demand for the remanufactured product at retailer s and is given by: Qrems,t = Qremt ωs where Qremt = qr∗ marketdemandt Decision variables: Iremt : inventory of remanufactured product at the beginning of a period Irett : inventory of returned product at the beginning of period t remcollecs,t : maximum number of product returns available at retailer s in period t zt : binary variable that takes a value of 1 if there is a shortage of remanufactured product in period t shortaget : quantity of shortage of remanufactured product in period t OpenDCdc : binary variable that takes a value of 1 if DC at d is open and 0 otherwise OpenHFhf : binary variable that takes a value of 1 if HF at hf is open and 0 otherwise OpenRCrc: binary variable that takes a value of 1 if RC at rc is open and 0 otherwise Qnewtr1mp,dc,s,t: quantity of new products, transported from manufacturing plant mp, through distribution center dc, to retailer s during period t Qnewtr2mp,hf,s,t: quantity of new products, transported from manufacturing plant mp, through hybrid facility hf, to retailer s during period t Qremtr1mp,dc,s,t: quantity of remanufactured products, transported from manufacturing plant mp, through distribution center dc, to retailer s during period t Qremtr2mp,hf,s,t: quantity of remanufactured products, transported from manufacturing plant mp, through hybrid facility hf, to retailer s during period t Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 75 Qrettr2s,hf,mp,t: quantity of consumer returns, transported from retailer s, through hybrid facility hf, to manufacturing plant mp during period t Qrettr1s,rc,mp,t: quantity of consumer returns, transported from retailer s, through recovery center rc, to manufacturing plant mp during period t Objective function: The objective function consists of the following cost components: • Revenue from sales (REV) = REVN + REVR1 where: – Revenue from new products (REVN) = (pn − cn ) N (1) Qnews,t t =1 sεR (i) For t < intro, we use Eq. (2) for pn and Eq. (3) for Qnewt , given below. p = n ∗ pn1 1 cn + = 2 2 λβ2 (α1 + β1 ) + (1 − λ) β1 (α2 + β2 ) λβ2 + (1 − λ) β1 ∗ marketdemandt Qnewt = qn1 (2) (3) α2 + β2 − pn α1 + β1 − pn + (1 − λ) Qnewt = λ marketdemandt β1 β2 (4) Qnews,t = Qnewt ωs (5) (ii) For t ≥ intro, we use Eq. (6) for pn and Eq. (7) for Qnewt , as shown below. pn = ∗ pn2 α1 + β1 cn + = max α1 , 2 2 ∗ (marketdemand ) λ = Qnew t= qn2 t λ2 min α1 +β1 −cn ,1 2β1 marketdemand t (6) (7) – Revenue from remanufactured products: ∗ Periods when we meet demand for remanufactured products, we use Eq. (9) for pr and Eq. (10) for Qremt , shown below (REVR1): (pr − cr ) N t =1 sεR Qrems,t − shortaget (8) 76 A. Dhamodharan and A. Ravi Ravindran where: α1 + β1 cr pr = max α2 , + 2 2 α2 + β2 − cr Qremt = λ (1 − λ) min , 1 marketdemandt 2β2 (9) Qrems,t = Qremt ωs (10) (11) • Inventory holding cost (IHC): N hr I remt + hc I rett (12) t =1 • Shortage cost (SHORT): N sc shortaget (13) (1 − ϕ) returnst dispcost (14) t =1 • Disposal cost (DISP): N t =1 • Transportation cost (TRANS): N t =1 mp ε MP dc ε DC s ε R N Qnewtr1mp,dc,s,t tf mp,dc,s + t =1 mp ε MP hf ε H F s ε R N t =1 mp ε MP dc ε DC s ε R N Qnewtr1mp,dc,s,t tf mp,dc,s + (15) t =1 mp ε MP hf ε H F s ε R N t =1 s ε R rc ε RC mp ε MP N Qnewtr1mp,hf,s,t tf mp,hf,s + Qnewtr1mp,hf,s,t tf mp,hf,s + Qnewtr1s,rc,mp,t tf s,rc,mp + t =1 s ε R hf ε H F mp ε MP Qnewtr1s,hf,mp,t tf s,hf,mp Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 77 • Fixed cost of opening facilities (FAC): fdc OpenDCdc + dc ε DC fhf OpenH Fhf + hf ε H F frc OpenRCrc rc ε RC (16) The objective function is to maximize the net profit: maximize: REV(IHC + SHORT+DISP+TRANS+FAC) Constraints: 1. Assign a route to all new products produced in period t to reach consumers through a DC or a HF. Qnewtr1mp,dc,s,t + mp ε MP dc ε DC Qnewtr2mp,hf,s,t mp ε MP dc ε DC = Qnews,t ∀s ε R, t ε T (17) 2. Assign a route to all remanufactured products produced in period t to reach consumers through a DC or a HF. Qremtr1mp,dc,s,t + mp ε MP dc ε DC Qremtr2mp,hf,s,t mp ε MP dc ε DC = Qrems,t − ωs shortaget ∀s ε R, t ε T (18) 3. Assign a route to all consumer returns collected in period t to reach MP through a RC or a HF. Qrettr1s,hf,mp,t + Qrettr2s,rc,mp,t hf ε H F mp ε MP rc ε RC mp ε MP = remcollecs,t ∀s ε R, t ε T (19) 4. Capacity constraint for each DC where the sum of all products transported to a DC in every period should not exceed its capacity, if the DC is built. Qnewtr1mp,dc,s,t + Qremtr1mp,dc,s,t mp ε MP dc ε DC ≤ capacitydc OpenDCdc ∀dc ε DC, t ε T (20) 78 A. Dhamodharan and A. Ravi Ravindran 5. Capacity constraint for each HF where the sum of all products transported to a HF in every period should not exceed its capacity, if the HF is built. Qnewtr2mp,hf,s,t + Qremtr2mp,hf,s,t + Qrettr2s,hf,mp,t mp ε MP s ε R ≤ capacityhf OpenH F hf ∀hf ε H F, t ε T (21) 6. Capacity constraint for each RC, where the sum of all consumer returns transported to a RC in every period should not exceed its capacity, if the RC is built. Qrettr1mp,rc,s,t + ≤ capacityrc OpenRC rc ∀rc ε RC, t ε T mp ε MP s ε R (22) 7. Constraint to update the inventory of consumer returns. Until the remanufactured product is introduced in the market, there is no production of remanufactured product. Hence, the inventory of consumer returns keeps adding up with the collection of returns. I rett = I rett −1 + remcollecs,t −1 ∀t < intro&t ε T (23) sεR 8. Constraint to update the inventory of consumer returns. When a remanufactured product is introduced in the market, there is a production of a remanufactured product. Hence, inventory at the start of the period is the sum of inventory from the previous period and consumer returns minus the quantity of returns used to produce the remanufactured product. I rett = I rett −1 − Qremt −1 + shortaget −1 remcollecs,t −1 ∀t < intro&t ε T sεR (24) 9. Quantity of consumer returns collected in period t that can be used to remanufacture is the product of remanufacturability and the parameter returnst . remcollecs,t ≤ ϕ γr returnst ∀s ε R, t ε T (25) Note: returnst is given by Equation in Appendix 2. 10. Constraint to update the inventory of remanufactured products. I remt = I remt −1 + ⎛ Qremt −1 − shortaget −1 ⎝ − Qremtr1mp,dc,s,t −1 + mp ε MP s ε R dc ε DC ⎞ Qremtr2mp,hf,s,t −1 ⎠ ∀t hf ε H F (26) Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 79 11. Constraint to detect shortage of remanufactured product in period t. If there is no shortage, zt = 0, else zt = 1. shortaget ≤ M zt ∀t (27) where M is a large number. 12. Non-negativity and binary constraints: I remt , I rett , Qremt , remcollect , shortaget ≥ 0∀t Qnewtr1dc,t , Qrem1dc,t ≥ 0∀dc, ∀t Qnewtr2hf,t , Qremtr2hf,t , Qrettr2hf,t ≥ 0∀hf, ∀t Qrettr1rc,t ≥ 0∀rc, ∀t (28) (29) (30) (31) OpenDCdc ε {0, 1} ∀dc, Open H Fhf ε {0, 1} ∀hf, OpenRCrc ε {0, 1} ∀rc (32) zt ε {0, 1} ∀t (33) 3.5 Case Study Applying PLCOM to Design Supply Chain Network for iPhone Problem Description We consider a CLSC with market segmentation consisting of quality-sensitive and price-sensitive consumers. The aim of the example is to illustrate the impact of considering the pricing model and the extended Bass model on the optimal supply chain network structure. Assuming each time period in PLCOM is 1 month long, we limit the lifecycle of the product to 100 periods (approximately 8 years) for better exposition. We assume the lifetime of a new product to be at least 12 periods (1 year) and up to 36 periods (3 years). Thus, a new product can only fail after 1 year of residence with the customer, and all products fail after 3 years (estimation by Apple). These are practical assumptions because electronic products are typically under warranty for the first 12 months and are not designed to last longer than 3 years. Hence, the residence time distribution ranges from period 12 to period 36. We use a symmetrical beta distribution with parameter 2.5 to generate the probability density 80 A. Dhamodharan and A. Ravi Ravindran Probability density functionResidence time 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 Time Periods Residence time distribution Fig. 3 Probability density function for residence time from periods 13 to 36 function for residence time, as shown in Fig. 3. We also assume that remanufactured products are introduced into the market in period 25. Based on the lifecycle assumptions above and considering iPhones to be a fast-moving product, we arrive at the innovative coefficient, innov = 0.0005 and imitation coefficient, imitation = 0.000025, so that the sales in the initial period of the lifecycle are significantly higher than the sales at the latter stage and the market consumes the product at least once within the product lifetime. We assume a realistic market structure with two groups of consumers: (i) quality-sensitive consumers with WTP function 550 + 250θ and (ii) price sensitive consumers with WTP function 450 + 200θ . We obtain the WTP function based on the iPhone market prices at different retailers. For quality-sensitive customers, we model the WTP function based on the iPhone7 (32 GB)‘s market price of $699, from the retailers like Amazon and Walmart. For price-sensitive customers, we model the WTP function based on the pre-owned iPhone6 Plus (16 GB) market price of $540, from the retailer Bestbuy. We assume the new product manufacturing cost to be $400 and the remanufacturing cost to be $100 (total price of all the components inside an iPhone is roughly estimated at $220). We assume 60% of the market is of quality-sensitive consumers (λ = 0.6) and 80% of the customers make a repeat purchase. We also assume all the collected returns are of good quality and can be remanufactured. We consider a CLSC network with 3 manufacturing plants, 3 retailers and 10 potential locations for distribution centers (DCs), recovery centers (RCs) and hybrid facilities (HFs), each with a capacity of processing 100 items. Recall that DCs handle only forward flows, RCs handle only reverse flows and HFs handle both forward and reverse flows. We assume the fixed cost of opening a DC and a RC to be $500,000, and the fixed cost of opening a HF to be $550,000. We assume per period inventory holding cost for a unit of new product, returned product and remanufactured product to be $5. Table 1 shows the proportion of market demand and product returns at each retailer. Appendix 4 lists unit transportation cost within the CLSC network. Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 81 Table 1 Proportion of market demand and product returns at each retailer Retailer 1 2 3 Proportion of market demand 0.3 0.4 0.3 Proportion of product returns 0.3 0.4 0.3 Problem Size The optimization problem has 30 binary variables, 48,066 continuous variables and 3972 constraints. The binary variables represent the selection of potential sites for DC, HF or RC. The continuous variables represent the distribution plan between manufacturing plants, DCs, HFs, RCs and retailers. The various constraints represent capacity and demand restrictions and conservation of flow at each facility for each period. The optimization problem was solved using CPLEX. The solver took 0.1619 seconds to solve the problem. Solution and discussion The maximum profit obtained by the CPLEX solver was $1,277,540. The optimal network structure uses 3 DCs (locations 1, 7 and 10) and 1 HF (location 10). No RC is built. Both a DC and a HF are built at location 10 since transportation cost to retailer 2 from location 10 is minimum. Retailer 2 has the maximum proportion of demand and the maximum proportion of returns (see Table 1). The optimal network flow for new products, remanufactured products and consumer returns are shown in Figs. 4, 5 and 6. In Fig. 4, we observe that the DC at location 1 serves new products to retailer 1, the DC at location 10 serves new products to retailer 2 and the DC at location 7 serves new products to retailer 3. In Fig. 5 we observe that DC at location 1 serves remanufactured product to retailer 1 and retailer 3, the DC and the HF at Fig. 4 Optimal network flow for new products 82 A. Dhamodharan and A. Ravi Ravindran Fig. 5 Optimal network flow for remanufactured products Fig. 6 Optimal network flow for consumer returns location 10 serves remanufactured product to retailer 2, the DC at location 7 serves remanufactured product to retailer 2 and retailer 3. In Fig. 6 we observe that the HF at location 10 collects all consumer returns from retailers 1, 2 and 3 and returns them to manufacturing plants MP1, MP2 and MP3. The flexibility offered by the HF in handling products both in forward and reverse supply chains is utilized by the model for supply chain efficiency. Figure 7 shows the evolution of product sales over the product lifecycle with respect to new products and remanufactured products. We note a decrease in new product sales in period 25, which is attributed to the introduction of the remanufactured product. Before the introduction of the remanufactured product, new products are sold to 73% of the market at $565 per phone. After the introduction of the remanufactured products, the new products are sold to only 60% of the market, which is the proportion of quality-sensitive consumers in the market, at $600 per phone. Thus, after the introduction of the remanufactured product, it is only optimal to sell the new products to quality-sensitive consumers and the demand for new products have been cannibalized by 13%. Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 83 Fig. 7 Evolution of new and remanufactured product sales Since we limit the number of times a customer can repeat the purchase of a product to one, we see that new product sales has a bimodal shape. The first peak in the new product sales curve is caused by the first-time customers, both qualitysensitive and price-sensitive consumers. The second peak is caused by the qualitysensitive repeat consumers. Hence, the cumulative new product sales volume before the new product sales peak is lower than the cumulative new product sales volume after the peak. If the spread of the residence time is narrower, we can see a more pronounced second peak due to repeat purchase, as discussed in Debo et al. (2006). The remanufactured product sales have a bell-shaped curve with a sharp increase to the peak, followed by a smooth decrease. The sharp increase to the peak is caused by the price-sensitive first-time consumers. The smooth decrease is due to the pricesensitive repeat consumers who replaced their first-time purchase of a new product with a remanufactured product. Figure 8 shows the evolution of product sales over the product life cycle with respect to first-time sales and repeat sales. We see the decrease in first-time sales at period 25, which can be attributed to the introduction of the remanufactured product into the market. We also observe that the peak of first-time sales occurs in period 43, and the peak of repeat sales occurs in period 68. The gap of 25 time periods between the peaks is explained by the unimodal shape of the residence function, with its maximum value at 25 periods after the purchase (see Fig. 3). The repeat sales is more skewed than the first time sales due to the delayed introduction of the remanufactured product. Figure 9 shows the availability of returns to remanufacture and shortages in remanufactured products. In period 25, we see a sharp increase in shortages as 84 A. Dhamodharan and A. Ravi Ravindran Evolution of First-Time Sales and Repeat Sales 350 300 250 200 150 100 Repeat Sales First-Time Sales Fig. 8 Sales across product life cycle: first-time sales and repeat sales 160 140 120 100 Fig. 9 Availability of returns and shortage of remanufactured product Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 85 sufficient returns are not available in the market to satisfy the demand for the remanufactured product. However, from period 57 onwards, we see no shortages of remanufactured products. 3.6 Illustration of Profit Reduction due to CLSC Network Design in Sequential Manner In this section, we illustrate to OEMs, the advantage of considering an integrated CLSC network design using PLCOM when initially setting up their supply chain network for a new product. PLCOM determines the optimal CLSC network design in an integrated manner, considering both forward and reverse supply chains together. We use the case study described previously to show the decrease in profit over the product life cycle, if the OEM solves the CLSC network design problem in a sequential manner instead of using PLCOM. In the sequential manner of determining the CLSC network design, the OEM initially determines the optimal forward supply chain network design and implements it. Later, when the OEM decides to participate in the remanufactured market, the OEM revisits the network design to obtain the optimal CLSC for the reverse supply chain. In the illustrative example, the OEM would first solve for optimal forward supply chain network design for the first 24 time periods (2 years) and implement the solution. Then, the OEM would solve for optimal CLSC network design from period 25 onwards using PLCOM, with the existing forward network design imposed as additional constraints to the PLCOM. The optimal forward supply chain network for the illustrative example consists of DCs at locations 1, 7, 9 and 10. Hence, we add additional constraints that DCs 1, 7, 9 and 10 are open to the PLCOM and solve for the optimal CLSC network when considering reverse supply chains also. The optimal CLSC network design adds the RC at location 7 to the optimal forward network design to accommodate the reverse flows. Table 2 compares the optimal network designs and profits obtained from the integrated and sequential approaches for CLSC network design. There is an additional RC chosen at location 7 and a DC at location 9 in the sequential approach, compared to the integrated approach. Also, no HF is chosen in the sequential approach. Since HFs were not chosen as part of the forward supply chain network design, at least one HF or RC had to be chosen to handle consumer returns. Also, the sequential approach imposes additional constraints to the PLCOM, resulting in a decrease in the profit compared to the integrated approach by 28%. Thus, if the OEM designs the CLSC network in a sequential manner, the profit over the product lifecycle will be significantly reduced due to the existing forward supply chain network. Hence, OEMs should take a long term view of the network design problem and consider both the forward and reverse flows in the network design as part of their decision-making process while launching a new product in the market. 86 A. Dhamodharan and A. Ravi Ravindran Table 2 Comparison of optimal network designs and profit between integrated and sequential approaches Factors Optimal network design Profit during product lifecycle Integrated CLSC DC:1,7,10; HF:10 $1,277,540 Sequential CLSC DC:1,7,9,10; RC:7 $923,222 3.7 Sensitivity Analysis Optimal Time Period to Introduce Remanufactured Product in the Market In this subsection, we conduct a sensitivity analysis on the parameter intro, the time period when the remanufactured product is introduced into the market, using the illustrative example. The value of the parameter intro is varied from 25 to 75, and the PLCOM is solved in an integrated manner to study its impact on the following: • • • • • • • network design, profit, shortage, new product sales, remanufactured product sales, first time sales and repeat sales. Impact on Profits Figure 10 plots optimal profit and shortage of remanufactured products, over the product lifecycle, against the intro values from 25 to 75. The optimal profit increases as the intro value increases to 45, then decreases smoothly until the intro value of 55. From the intro values of 60 onwards, we see a sharp decrease in the optimal profit. We conclude that 40 is the most profitable time period to introduce the remanufactured product into the market. Time period 60 and beyond is too late for the OEM to realize maximum profits from the remanufacturing market. Impact on Shortages Also note that in Fig. 10, the total shortage over the product life cycle is 3500 for the intro value of 25. As the intro value increases, that is, when the remanufactured products enter the market later, there is a decrease in shortages because there are more product returns available for remanufacturing. Interestingly, the minimum shortage value of 0 occurs for intro values of 55 or more. The maximum profit scenario with intro value of 40 has a shortage of 744 remanufactured units. Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 87 Fig. 10 Profit and shortage over product life cycle Impact on New and Remanufactured Product Sales Figure 11 plots new product sales, remanufactured product sales and total sales over the product lifecycle against the intro values from 25 to 75. We observe that both remanufactured product sales and the total sales decrease in a linear fashion with higher values of intro. This is explained by the fact that as the value of Fig. 11 New and remanufactured product sales over product life cycle 88 A. Dhamodharan and A. Ravi Ravindran Fig. 12 First time and repeat sales over product life cycle intro increases, the time periods left for the price-sensitive consumers to purchase the remanufactured products reduce. On the other hand, we see an increase in new product sales for higher values of intro because the new product demand is cannibalized after intro. Impact on First-Time and Repeat Sales Figure 12 plots first time sales repeat sales and total sales, over the product lifecycle against the intro values from 25 to 75. We observe that first-time sales decreases with an increase in intro value, while the repeat purchase remains constant. Firsttime sales made by the price-sensitive and the quality-sensitive consumers decreases with an increase in intro value since price-sensitive consumers do not have enough time to make first-time purchases of remanufactured products. There are only a fixed number of quality-sensitive consumers in the market. The model restricts the number of repeat purchases to one. Hence, repeat purchases are flat and independent of intro value. 3.8 Sensitivity Analysis Impact on the Supply Chain Network Table 3 shows the optimal CLSC network design for different values of intro. It is interesting to note that the DCs and the HFs chosen are the same irrespective of the Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design Table 3 Optimal network designs for different values of intro Intro 25 30 35 40 45 50 55 60 65 70 75 DC 1,7,10 1,7,10 1,7,10 1,7,10 1,7,10 1,7,10 1,7,10 1,7,10 1,7,10 1,7,10 1,7,10 HF 10 10 10 10 10 10 10 10 10 10 10 89 RC – 10 1 10 7 – – – – – – intro value. As intro value is increased from 25 to 30, 35, 40 and 45, we find that it is profitable for the CLSC network to operate an additional RC in order to reduce the shortage of the remanufactured products. For intro value of 50, in spite of a shortage of 534 units of the remanufactured products (see Fig. 11), it is not profitable to operate the additional RC as the fixed cost of the facility exceeds the profit made from the sales of the 534 units. For intro value of 55 onwards, it is not profitable to operate the additional RC due to the reduced volume of the remanufactured sales. In summary, for the case study, the OEM can achieve a maximum profit of $1,514,000 by introducing the remanufactured product at time period 40. The optimal CLSC network will have: • DCs at locations 1, 7 and 10 • an RC at location 10 • a HF at location 10 In other words, just at location 10, the OEM will have a DC, RC and HF. The reason behind choosing location 10 is the minimum transportation cost to retailer 2, who faces the maximum proportion of demand and collects the maximum proportion of returns (see Table 1). 3.9 Characterization of Slow and Fast Diffusing Products In this subsection, we conduct a sensitivity analysis on the parameter imitation, which creates new demand through the word-of-mouth effect. As the value of the imitation coefficient is increased, the produce lifecycle grows shorter. We define fast diffusing products as products with lifecycle shorter than twice the maximum product lifetime and slow diffusing products as products with lifecycle longer than twice the maximum product lifetime. In our case study, the maximum product lifetime is 36 time periods. Hence, products with a lifecycle shorter than 72 periods are considered fast diffusing products, and products with a lifecycle longer than 72 90 A. Dhamodharan and A. Ravi Ravindran periods are considered slow diffusing products. We solve the illustrative example using the PLCOM for lifecycle values of 100 periods, 80 periods, 60 periods and 40 periods. The products with lifecycle values of 40 and 60 periods are considered as fast diffusing products, while the products with lifecycle values of 80 and 100 periods are considered as slow diffusing products. Note that in all the 4 cases the total market size is maintained at 10,000. Impact on Profit and Network Design Table 4 compares the profit and the optimal network designs for the abovementioned values of the product lifecycle. We find that the fast diffusing product with a lifecycle value of 40 periods results in a loss to the OEM, as it requires extensive investment to open DCs at 6 different locations and HF at location 10, in order to handle the high volume of product flow in each period. Hence, when designing a supply chain for fast diffusing products, we recommend the OEM to plan for flexible capacity, so that the fixed cost of opening facilities is minimal, as noted by Debo et al. (2006). In the case of a fast diffusing product with a lifecycle value of 60 periods, the OEM is able to make a profit by opening DCs at 3 different locations and HF at location 10. The profit is significantly lower than the profit made from slow diffusing products due to a high shortage of consumer returns. In the case of slow diffusing products, we notice that the profit is higher for a product with a longer lifecycle. Under a fixed capacity setting, as the product lifecycle grows longer, the demand is more evenly distributed. Thus, we expect the number of facilities in the supply chain network to remain about the same. From Table 4, we notice that the number of facilities in the network for the lifecycle value of 60 periods is lower than the number of facilities for the lifecycle value of 40 periods. Similarly, the number of facilities in the network for the lifecycle value of 80 periods is lower than the number of facilities for the lifecycle value of 40 periods. Finally, the number of facilities in the network for the lifecycle value of 100 periods is lower than the number of facilities for the lifecycle value of 40 periods and 80 periods and is the same as the number of facilities in the network for the lifecycle value of 60 periods. Interestingly, we observe that the number of facilities in the network for the lifecycle value of 80 has an additional facility compared to the network for the lifecycle value of 60. The additional facility is to satisfy the demand for the remanufactured products from cannibalization, explained later in this section. Table 4 Profit and optimal network design for slow and fast moving products Category Fast diffusing Fast diffusing Slow diffusing Slow diffusing Lifecycle 40 60 80 100 Profit −$1,853,000 $390,900 $1,035,000 $1,277,540 DC 1,2,7,8,9,10 1,7,10 1,7,10 1,7,10 HF 10 10 10 10 RC – – 7 – Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 91 Impact on New and Remanufactured Product Sales Figure 13 plots new product sales and remanufactured product sales over the product lifecycle for different lifecycle values. We notice that as the product lifecycle shortens, the peaks from the first-time sales and the repeat sales are more pronounced. In the case of a fast diffusing product with a lifecycle value of 40 periods, we see a big drop in the new product sales when the remanufactured product is introduced. This indicates a big proportion of new product consumers switching from the new product to buy the remanufactured product. The revenue loss from the demand cannibalization also results in a net loss to the OEM, as shown in Table 4. Thus, in this case, it is clearly not profitable for the OEM to participate in the remanufacturing market. In the case of fast diffusing products with a lifecycle value of 60 periods, we notice that the demand cannibalization is significantly lower, and the OEM makes a net profit. In order to analyze whether it is profitable for the OEM to participate in the remanufactured market, we solve only for the forward supply chain and find that the OEM would incur a net loss operating only in the forward supply chain, due to fixed capacity. Hence, we conclude that in this case, it is profitable for the OEM to participate in the remanufacturing market. Fig. 13 New and remanufactured product sales for different values of product lifecycle 92 A. Dhamodharan and A. Ravi Ravindran Impact on First-Time Sales, Repeat Sales and Shortage Figure 14 plots first-time sales, repeat sales and shortage over the product lifecycle, for different lifecycle values. We notice that, as the product lifecycle grows shorter, there is an increase in the amplitudes of the first-time sales, repeat sales and the shortage functions. We also find that the total first-time sales increases and the total repeat sales decreases, due to lack of time for the product to fail. Initially, the total shortage increases due to the shortage of consumer returns. In the case of lifecycle value of 60 periods, we notice a second peak in shortage, coinciding with the peak of repeat sales. This second peak is due to increased demand for remanufactured products from the delayed cannibalization effect. The delayedcannibalization effect occurs when the price-sensitive consumers, who initially bought the new product, before the introduction of the remanufactured product, switch to buy the remanufactured product when making their repeat purchase. In the illustrated example, the delayed cannibalization effect is expected to occur from periods 38 to 61, as shown in Fig. 14. As the product lifecycle grows longer, the number of consumers causing the delayed cannibalization effect is smaller and hence more evenly distributed, providing time to collect consumer returns and satisfy the increased demand for the remanufactured product. Thus, in the case of the lifecycle value of 80 periods, we observe the second shortage peak with a significantly lower amplitude, and in the case of the lifecycle value of 100 periods, there is no second shortage peak. This also explains why the optimal network design for the lifecycle value of 80 periods has an additional RC at location 7 compared to the lifecycle value of 60 periods. Fig. 14 First-time sales, repeat sales and shortage for different values of product lifecycle Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 93 The additional RC is used to collect more returns and minimize the shortage, as illustrated in Fig. 14. However, the additional RC is not required in the case of the lifecycle value of 100 periods, because of the very small increase in the demand for remanufactured products from the delayed cannibalization. In summary, we conclude the following from the sensitivity analysis of the imitation coefficient: • Slow diffusing products are more profitable for remanufacturing compared to fast diffusing products because there is enough time to collect consumer returns • OEMs dealing with fast diffusing products should plan for flexibility in capacity in order to avoid the fixed cost of opening additional facilities • Managers should understand the effect of delayed cannibalization when designing the reverse channel and planning for the collection of consumer returns 3.10 Incentivizing Consumers to Improve the Quality of Returns We conduct a sensitivity analysis on the quality of the consumer returns, which is defined as the percentage of consumer returns that are qualified to be remanufactured. In the illustrative example, we assume that 80% of the returns are qualified to be remanufactured. We solve the illustrative example using the PLCOM for consumer return quality values of 60%, 70%, 80% and 90%. Table 5 shows the profit and the optimal network design for the 4 values of quality mentioned above. We expect that as the quality of the consumer returns increases, the profit should also increase, as shown in Table 5. Also, we notice from the table that the number of facilities in the optimal network increases with an increase in the quality of the returns. This is also expected as an increase in the quality of the consumer returns increases the available remanufactured products, resulting in the handling of a larger number of consumer returns and remanufactured products by the CLSC network. Note that there are 4 facilities in the optimal network for the quality values of 60%, 70% and 80% while there are 5 facilities in the optimal network for the quality value of 90%. The additional RC facility handles the extra remanufactured products produced from the higher-quality consumer returns. Table 5 Profit and optimal network design for different values of the quality of consumer returns Quality of consumer returns 60% 70% 80% 90% Profit $960,000 $1,190,000 $1,277,540 $1,454,000 DC 7,10 1,7,10 1,7,10 1,7,10 HF 1 – 10 10 RC 7 10 – 1 94 A. Dhamodharan and A. Ravi Ravindran It is in the interest of the OEMs to collect consumer returns of high quality. Thus, the OEMs can provide incentives to consumers to encourage high-quality returns. Such a practice is followed in the smartphone industry, where the consumer gets a rebate when buying a new smartphone for returning a good quality smartphone. The difference in profits for the various quality values (Table 5) provides an upper bound on the incentives that can be provided by the OEM, without incurring loss. For example, if the OEM is collecting consumer returns of 60% quality, then the OEM is justified to spend up to $229,700 (the difference between $1,1,90,000 and $960,300) in consumer incentives, to ensure a 70% quality of consumer returns. Figure 15 plots the new product sales and the remanufactured product sales over the product lifecycle, for different values of the quality of consumer returns. We notice that, as the quality of the consumer returns increases, the second peak of the new product sales and the remanufactured product sales increase due to an increase in repeat purchase. We also find that as the quality of the consumer return increases, the total new product sales and the total remanufactured product sales increase in a linear fashion. Figure 16 plots the first-time sales, the repeat sales and the shortage over the product lifecycle, for different values of quality of consumer returns. We notice that as the quality of the consumer returns increases, the total first-time sales and the total repeat sales increase. However, as the quality of consumer returns increases, the total shortages do not monotonically increase. This can be explained by the following example: When the quality of consumer return value increases from 60% from 70%, the total shortage increases from 2743 units to 3079 units. The profit in the case of 70% quality of consumer return is still higher than the profit in the case Fig. 15 New and remanufactured product sales for different values of the quality of consumer returns Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 95 Fig. 16 First-time sales, repeat sales and shortage for different values of the quality of consumer returns of 60% quality of consumer returns (see Table 5) due to the cost savings from the network design. Note that the optimal network design for 60% quality of consumer returns opened a HF at location 1, which costs $50,000 more than the DC at location 1 opened in the optimal network design for 70% quality of consumer returns (see Table 5). Since a HF can handle both products and returns but a DC can only handle products, we notice an increase in shortage in the case of 70% quality of consumer returns. In summary, we conclude the following from the sensitivity analysis on the quality of the consumer returns: • Higher quality of consumer returns provide higher profit for the OEMs • The analysis provides OEMs insights on how much to spend on consumer incentives in order to improve the quality of the consumer returns, without incurring a net loss from the incentives Total sales usually increase with an increase in the quality of consumer return • Total shortages do not always decrease with an increase in the quality of consumer returns. Sometimes it may be optimal to have a higher shortage due to the high fixed cost of opening an additional facility. 96 A. Dhamodharan and A. Ravi Ravindran 4 Conclusions In this chapter, we began by introducing the PLCOM and illustrated the evolution of new and remanufactured product sales, first-time and repeat sales and shortage of remanufactured products over the product lifecycle using a case study. The case study also quantified the cannibalization of the new product demand, a phenomenon that deters the OEMs from participating in the market for the remanufactured products. We continued with the same example to illustrate that an integrated approach to design a CLSC network is more profitable for the OEM to participate in the reverse supply chain. Sensitivity analysis on the time period to introduce the remanufactured product in the market identified the optimal time period maximizing the profit and showed the tradeoff between the shortage of remanufactured products and the new product sales. Sensitivity analysis on imitation coefficient showed that slow diffusing products are more profitable in the reverse channel compared to fast diffusing products. We also identified the delayed cannibalization phenomenon in case of fast diffusing products caused by the price-sensitive consumers switching to the remanufactured product during their repeat purchase. Finally, the sensitivity analysis on the quality of consumer returns provided insights into how much the OEM can spend on incentives to improve the quality of consumer returns, without incurring a net loss from the incentives. Appendices Appendix 1 Pricing Model Quality-sensitive consumers will only buy new products and are not motivated by price discounts. The willingness to pay (WTP) function for quality-sensitive customer segment is modeled as a linear function given by: f (θ ) = α1 + β1 θ (34) where θ in (0,1) and (1- θ) is the fraction of the customer segment with WTP = α1 + β1 θ. Price-sensitive consumers are indifferent to whether the product is new or remanufactured and are purely motivated by the price of the product. Hence, there is no need to model new and remanufactured products separately. The willingness to pay (WTP) function for this customer segment is modeled as a linear function given by: f (θ ) = α2 + β2 θ (35) Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 97 where θ in (0,1) and (1- θ) is the fraction of the customer segment with WTP = α2 + β2 θ, α2 is the maximum selling price at which all the customers in this price-sensitive category are willing to buy the product (new or remanufactured). The proportion of price-sensitive consumers, who are willing to buy new/remanufactured product for price pr , denoted by qr , is given by: pr − α2 qr = (1 − λ) 1 − β2 (36) where λ is the fraction of quality-sensitive customers. ∗ pn1 = 1 λ β2 (α1 + β1 ) + (1 − λ) β1 (α2 + β2 ) cn + 2 2 λ β2 + (1 − λ) β1 (37) λ β2 (α1 + β1 ) + (1 − λ) β1 (α2 + β2 ) λ β2 + (1 − λ) β1 (38) ∗ qn1 = where: cn : manufacturing cost of new product cr : remanufacturing cost λ: the fraction of quality-sensitive consumer segment 1- λ: the fraction of the market falling under price-sensitive consumer segment In order to make use of the above equations, we assume that α2 < α1 ≤ α2 + β2 , ∗ ≥ i.e., there is overlapping in WTP between the two consumer categories. Also, pn1 ∗ α1 and α2 ≤ pn1 ≤ α2 + β2 . When the OEM introduces the remanufactured product into the market, assuming no constraint on the production of remanufactured products, the optimal pricing pn∗ , pr∗ and the optimal production quantity qn∗ , qr∗ , expressed as a fraction of the market demand for new products and remanufactured products, respectively, are given by the following equations (Abbey et al. 2015a). α1 + β1 cn ∗ + = max α1 , pn2 2 2 ∗ qn2 = λ min α1 + β1 − cn ,1 2β1 (39) (40) α2 + β2 cr ∗ pr2 = max α2 , + 2 2 ∗ qr2 α2 + β2 − cr = (1 − λ) min ,1 2β2 (41) (42) We consider a multi-period optimization model in which the OEM tries to maximize the revenue in each time period. The OEM initially sells only the new 98 A. Dhamodharan and A. Ravi Ravindran product until the introduction of the remanufactured product. During those periods, we use Eqs. (37) and (38) to decide the pricing and quantity of the new product produced. Let intro be the time period when the OEM introduces the remanufactured product into the market. During the periods after the introduction of remanufactured products, assuming no capacity restriction on producing new products, we use Eqs. (39) and (40) as the new price and the new production quantity for the new product. For the remanufactured products, however, the optimization model constraints the production quantities based on the quantity of consumer returns and the remanufacturability of the consumer returns. Hence, if the availability of the consumer returns in a period exceeds the demand for remanufactured products, we use the Eqs. (41) and (42) to make the pricing and the production quantity decisions for the remanufactured products. If the availability of the consumer returns in a period is lower than the remanufactured product demand, then we assume that the OEM sells all the available inventory of remanufactured products at pr∗ , and the shortages are lost sales. In order to make use of the Eqs. (39) through (42), we need to make the following assumptions: cn > cr . The manufacturing cost of the new product is greater than the remanufacturing cost α1 > α2 . The minimum WTP for the new product is greater than the minimum WTP for the remanufactured product α1 + β1 > α2 + β2 . The maximum WTP for the new product is greater than the maximum WTP for the remanufactured product All the three assumptions are very practical. Under such assumptions, pn∗ > pr∗ . Hence, when remanufactured product is available in the market, pricesensitive customers will not buy the new product. Thus, new product demand might decrease due to the introduction of the remanufactured product, resulting in demand cannibalization. Appendix 2 Demand Model Debo et al. (2006) extended the Bass model to study life cycle demand of new and remanufactured products and analyzed the cost-effectiveness of investments in capacity for remanufacturing, as explained in the literature review section. Debo et al. (2006) assumed a specific OEM selling price function for both the new and the remanufactured products, without considering the market structure. We adopt their model to generate demand for both new and remanufactured products throughout the product life cycle and then integrate it with the OEM’s pricing model, which is described in the previous section. The demand model is described below. Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 99 Product Life Cycle Parameters: marketsize: refers to the total size of the market considered. marketsize is normalized to 1 marketdemandt: total demand (for new and remanufactured products) in period t as a fraction of the marketsize marketexpt : market size expansion for both new and remanufactured products due to the extended Bass model, in period t, expressed as a fraction of the total market size markett : market size of first-time product owners using new and remanufactured products at the end of period t, expressed as a fraction of the total market size innov: innovation coefficient of the product that creates new demand. The coefficient is expressed as a fraction of the total market size imitation: imitation coefficient that creates new demand through word-of-mouth effect from existing quality-sensitive consumers. This coefficient is also expressed as a fraction of total market size ϕ: the fraction of current product owners, making a repeat purchase. This parameter is treated as a constant over time residencej : the fraction of current product owners, whose products will fail at period j repeatt : demand from repeat purchases by current product owners in period t, expressed as a fraction of markett L: product lifetime intro: time period when OEM introduces new product newprdownt : number of consumers who own the new product in period t returnst : number of consumer returns of new product in period t. Note that we do not consider returns of the remanufactured product. The consumer return may be a commercial return or end of life return Qnewt : demand for new product at the beginning of period t. Since we assume no capacity restrictions at MP, this is also the number of consumers who bought the new product at the beginning of period t. Qnewt = qn∗ (marketdemandt) Demand Equations In the extended Bass model, the market size for new and remanufactured products grows each period due to the appeal of the product to the market, denoted by the innovation coefficient, and word-of-mouth from the existing customers, denoted by the imitation coefficient. Market size expanded during period t is a fraction of the uncaptured market (1-markett-1) that is open to buying the product, as given by the Eq. (43). marketexpt = (innov + imitation newprodownt −1 ) (1 − markett −1) (43) 100 A. Dhamodharan and A. Ravi Ravindran Market demand for new and remanufactured products in period t, given by the Eq. (44), is the sum of market expanded during period t and repeat consumers. marketdemandt = (marketexpt + repeatt ) marketsize (44) Market demand is split into new and remanufactured product demand using the parameter λ, defined as the fraction of quality-sensitive customers in the market. Hence, the new product demand is given by λ (marketdemandt) and the remanufactured product demand is given by (1 − λ) (marketdemandt). The market size of first-time product owners, given by Eq. (45) using both new and remanufactured products at the end of period t, is updated as the sum of the market size at the end of the previous period and the market size growth during period t. markett = markett −1 + marketexpt (45) In the extended Bass model, the term residence time distribution is used to refer to the probability distribution of the number of time periods for which the product (new or remanufactured) stayed with a customer. Hence, the number of new product owners at the beginning of period t is obtained by subtracting the number of new product consumers who stopped owning the product at the end of the period (t-1) and adding the number of consumers who bought the new product at the beginning of period t, as given by the Eq. (46). newprodownt = newprodownt −1 +Qnewt − L residencej newprodownt −j −1 j =1 (46) where Qnewt = qn∗ marketdemandt . At the end of the residence time, the customer is assumed to return the new product. At this time, the customer is treated as a potential repeat customer of the new product. Parameter repcust represents the fraction of the potential repeat customers who actually repeat purchase, making them repeat customers. The size of the repeat customers in period t is given by Eq. (47). repeatt = repcust min(t,L) residencej − marketexpt −j + repeatt −j (47) j =1 Consumer returns at the beginning of period t is given by the sum product of residence and market size that owns new products. Repeat consumers, part of newprodown, can once again return their new product, as given by Eq. (48). returnst = L j =1 residencej newprodownn,t −j (48) Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 101 Appendix 3 Integration of Pricing and Demand Models In the multiperiod optimization model, we model the demand (both new and remanufactured products) in each period based on the extended Bass model, given by Eqs. (44) to (46). However, the production quantities of new and remanufactured prod∗ , q ∗ , q ∗ ), are expressed as a fraction of the market demand for new products (qn1 n2 r ucts and remanufactured products, based on the pricing model, as given below: 1. During time periods when only the new products are produced (upto t < intro), ∗ we use Eq. (38) from Appendix 1 to determine qn1 2. During time periods when both new and remanufactured products are produced ∗ , for new t ≥ intro), we use Eq. (40) from Appendix 1 for determining qn2 products, 3. For remanufactured products, we use Eq. (42) from Appendix 1 for determining qr∗ , The marketdemand, for new and remanufactured products in period t is given by Eq. (44). Using marketdemand, the production quantities for new and remanufactured products are determined as follows: Let Qnewt be the production quantity of new product. Then, ∗ , we get: (i) Using Eq. (38) from Appendix 1 for qn1 Qnewt = λ α1 + β1 − pn∗ 1 − β1 + (1 − λ) α2 + β2 − pn∗ 1 − β2 marketdemandt (49) for period t < intro, when only the new product is produced. ∗ , we get: (ii) Using Eq. (40) from Appendix 1 for qn2 Qnewt = λ min α1 + β1 − cn ,1 2β1 marketdemandt (50) for period t, when both the new product and the remanufactured product are produced. (iii) Let Qremt be the weighted production quantity for the remanufactured product. Qremt = qr∗ marketdemandt . Using Eq. (42) from Appendix 1 for qr∗ , we get: Qremt = (1 − λ) min α2 + β2 − cr ,1 2β2 marketdemandt (51) 102 A. Dhamodharan and A. Ravi Ravindran Appendix 4 Table 6 Transportation cost from manufacturing plant 1 Plant 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 DC/HF/RC location 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 Retailer 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 Cost ($) 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5 0.7 0.75 0.8 0.85 0.9 0.95 0.5 0.55 0.6 0.65 Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 103 Table 7 Transportation cost from manufacturing plant 2 Plant 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 DC/HF/RC location 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 Retailer 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 Cost ($) 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5 0.7 0.75 0.8 0.85 0.9 0.95 0.5 0.55 0.6 0.65 Table 8 Transportation cost from manufacturing plant Plant 3 3 3 3 3 3 3 3 3 DC/HF/RC location 1 2 3 4 5 6 7 8 9 Retailer 1 1 1 1 1 1 1 1 1 Cost ($) 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 (continued) 104 Table 8 (continued) A. Dhamodharan and A. Ravi Ravindran Plant 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 DC/HF/RC location 10 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 Retailer 1 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 Cost ($) 0.95 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6 0.55 0.5 0.7 0.75 0.8 0.85 0.9 0.95 0.5 0.55 0.6 0.65 References Abbey, J. D., Meloy, M. G., Guide, V. D. R and Atalay, S. (2015aa), ‘Re-manufactured products in closed-loop supply chains for consumer goods’, Production and Operations Management 24(3), 488–503 Abbey, J. D., Blackburn, J.D. and Guide, V.D.R. (2015bb) ‘Optimal pricing for optimal pricing for new and remanufactured products’, Journal of Operations Management 24(3), 488–503. Bass, F. M. (2004). Comments on “a new product growth for model consumer durables the bass model”. Management Science, 50(12 supplement), 1833–1840. Chen, J.-M., & Chang, C.-I. (2013). Dynamic pricing for new and remanufactured products in a closed-loop supply chain. International Journal of Production Economics, 146(1), 153–160. Debo, L. G., Toktay, L. B., & Wassenhove, L. N. V. (2006). Joint lifecycle dynamics of new and remanufactured products. Production and Operations Management, 15(4), 498–513. Ferrer, G., & Swaminathan, J. M. (2010). Managing new and differentiated remanufactured products. European Journal of Operational Research, 203(2), 370–379. Guide, V. D. R., Jr., & Li, J. (2010). The potential for cannibalization of new products sales by remanufactured products. Decision Sciences, 41(3), 547–572. Jayaraman, V. (2006). Production planning for closed-loop supply chains with product recovery and reuse: An analytical approach. International Journal of Production Research, 44(5), 981– 998. Hammond, D., & Beullens, P. (2007). Closed-loop supply chain network equilibrium under legislation. European Journal of Operational Research, 183(2), 895–908. Majumder, P., & Groenevelt, H. (2001). Competition in remanufacturing. Production and Operations Management, 10(2), 125–141. Product Life Cycle Optimization Model for Closed Loop Supply Chain Network Design 105 Min, H., Ko, C. S., & Ko, H. J. (2006). The spatial and temporal consolidation of returned products in a closed-loop supply chain network. Computers & Industrial Engineering, 51(2), 309–320. Ovchinnikov, A. (2011). Revenue and cost management for remanufactured products. Production and Operations Management, 20(6), 824–840. Pang, G., Casalin, F., Papagiannidis, S., Muyldermans, L., & Tse, Y. K. (2015). Price determinants for remanufactured electronic products: A case study on ebay UK. International Journal of Production Research, 53(2), 572–589. Pazhani, S., & Ravindran, A. R. (2014). Design of closed loop supply chain networks. International Journal of Business Analytics, 1(1), 43–66. Savaskan, R. C., Bhattacharya, S., & Van Wassenhove, L. N. (2004). Closedloop supply chain models with product remanufacturing. Management Science, 50(2), 239–252. Toktay, L. B., & Wei, D. (2011). Cost allocation in manufacturing– Remanufacturing operations. Production and Operations Management, 20(6), 841–847. Supply Chain Risk Management in Indian Manufacturing Industries: An Empirical Study and a Fuzzy Approach V. Viswanath Shenoi, T. N. Srikantha Dath, and Chandrasekharan Rajendran 1 Background and Motivation The evolution of the free trade agreement during the past decade has facilitated the movement of goods across the world (Moore and Moore, 2003). This enabled companies to compete in the international markets with products produced for the domestic markets without any trade barriers. Further, the liberalization and the consequent economic reforms in India such as Foreign Direct Investment (FDI) led to substantial investments in the manufacturing sector which not only augmented as an opportunity for the revival of the Indian economy (Kumar, 2005) but also led to increased competition. Globalization and competitive environments necessitated companies to implement an aggressive and integrated enterprise-wide approach towards risk management (Saeidi et al., 2019). Hence the implementation of Supply Chain Risk Management, a critical strategic approach has become a top priority. For managing and mitigating risk, effective SCRM systems control adverse outcomes through systematic implementation are required (Manuj V. Viswanath Shenoi () Department of Computer Science and Engineering, Amrita College of Engineering and Technology, Nagercoil, Tamil Nadu, India e-mail: v_viswanathshenoi@amrita.edu.in T. N. Srikantha Dath Department of Mechanical and Manufacturing Engineering, M S Ramaiah University of Applied Sciences, Bengaluru, Karnataka, India e-mail: srikanthadath.me.et@msruas.ac.in C. Rajendran Department of Management Studies, Indian Institute of Technology Madras, Chennai, Tamil Nadu, India e-mail: craj@iitm.ac.in © Springer Nature Switzerland AG 2021 S. Srinivas et al. (eds.), Supply Chain Management in Manufacturing and Service Systems, International Series in Operations Research & Management Science 304, https://doi.org/10.1007/978-3-030-69265-0_4 107 108 V. Viswanath Shenoi et al. et al., 2014). This study provides a platform for researchers to study the concept of SCRM and discover the underlying issues. Further, the practicing managers can utilize the presented instruments to review and to redefine their SCRM issues. The study is done in two folds. First, the critical factors of SCRM are identified and a framework is developed. The developed framework helps the researchers to understand the relationship between the critical risk-related factors. The empirical study is performed to gain the perceptions of the practicing managers about risks perceived in their organization. In the second part, the critical factors of SCRM are constituted as the states of the Fuzzy Cognitive Map, which represents the dynamical system of the supply chain. The system helps researchers to identify all plausible risks in the long run, given a risk observed from a point of time, and suggests mitigation strategies for practicing managers. This chapter is organized as follows: Sect. 2 describes the literature evidence of the supply chain risk and mitigation strategies. Section 3 deals with the methodology of the research frameworks and explains each of the methods used in the study. Then results are presented in Sect. 4. The detailed discussion of the results can be seen in Sect. 5 and the conclusion in Sect. 6. 2 Literature Review 2.1 Critical Factors for SCRM According to the Supply Chain Operations Reference Model (SCOR) model, the possible disruptions in a supply chain may occur inside the supply chain, like inconsistent in quality or uncertainty in demand/supply, or external to the supply chain, namely, strikes, natural calamities or terrorist attacks. Hence, the probable disruptions should be systematically identified, evaluated and mitigated through SCRM. The supply chain disruptions arise from three primary sources, namely, environmental risks (e.g., socio-political actions and natural disasters), organizational risks (e.g. machine breakdown, labor strikes and IT break-down etc.), and network risks (e.g. interactions within the supply chain such as information sharing) (Sodhi and Chopra, 2004; Ghadge et al., 2013). Wagner and Bode (2008) categorized the risks based on their location of occurrence. Demand side and supply side risks are categorized as internal supply chain risks, and bureaucratic, infrastructure, and catastrophic risks are categorized as external supply chain risks. Trkman and McCormack (2009) classified risk based on their source i.e. within or external to the supply chain. Oke and Gopalakrishnan (2009) considered factors including the likelihood of a risk and its impact for classification. SCRM in Indian Manufacturing Industries 109 Description of the Construct of SCRM: Manufacturer’s Perspective Demand Side Risk1 (DSR) For good business performance, organizations need to deliver the required quantities of the product at the pre-agreed time and place. Issues such as volatile customer demand, inaccurate demand forecasting, competitive rival products, poor understanding of customer preferences, short product life cycle, and defects in the products are predominant demand side risks (Sodhi and Chopra, 2004; Wagner and Bode, 2006; Oke and Gopalakrishnan, 2009; Trkman and McCormack, 2009). Supply Side Risk1 (SSR) The supplier plays an important role in providing the sub-assemblies to the original equipment manufacturers (OEMs) or (Tier-2 supplier to Tier-1 supplier). The risks faced by the suppliers are due to market characteristics, quality problems, financial instability, technological changes, and product design issues (Wagner and Bode, 2006; Trkman and McCormack, 2009; Thun and Hoenig, 2011). Logistic Risk (LR) Logistics service providers deliver products and services at the predestined place at the appropriate time, which is a characteristic of an effective and efficient supply chain. For efficient logistics, choice of a transport, the type of vehicle, and their utilization play a major role (KPMG, 2010; Rogers et al., 2012). Regulatory, Legal and Bureaucratic risk1 (RLB) Government, top management and administration take part in important decisions pertaining to planning (longterm, short-term), regulatory policies and reforms.While implementing a new plan, policy, or reform, apart from confirming the overall cost-effectiveness, it has to be ensured that every stakeholder in the supply chain gets an advantage (Hendricks and Singhal, 2005; Wagner and Bode, 2006; Rogers et al., 2012). Infrastructure Risk1 (IR) The infrastructure of the firm includes humans, machines, buildings, and service systems, as they are all essential for the growth and sustenance of the firm. So, it is mandatory for the firm to maintain cordial Human Resource policies and perform regular maintenance (breakdown, preventive, corrective) to machinery (FCCI, 2013). Stock/Data Management risk1 (SDM) Stock information needs to be maintained and shared across the supply chains Specifically for Vendor Managed Inventory (VMI), it is vital to dispense the information across all the stakeholders of the supply chain (Dath et al., 2009). Firms needs to maintain standard product identification, EDI, CFAR, and customer preferences details to enable the vendors and suppliers to know the expansions in the stock levels and replenish them suitably. Environmental Risk1 (ER) Natural hazards, terrorist attacks on the establishment, civil unrest, and epidemics disrupting the supply chain have been reported from the 1 Indicates the factors adapted appropriately and suitably from the work of Sodhi and Chopra (2004) and Wagner and Bode (2006). 110 V. Viswanath Shenoi et al. various parts of the world. Certain events may be unavoidable, but one needs to be prepared to face unfavorable/undesirable events (Oke and Gopalakrishnan, 2009; Thun and Hoenig, 2011). Financial Risk1 (FR) Firms need to have sound financial stature to sustain in the competitive environment. Hence, it vital to maintain the optimum cash flow into the business from the customers on time. Also, financial activities such as diversification of the funds, inappropriate drawing of loans to fund the business, over-leveraging of the funds out of business in high-risk investments are undertaken to maximize the returns that may prove fatal to the firms’ existence (NMCC, 2008; Dath et al., 2009; CRISIL, 2010). Top Management Commitment2 (TMC) SCRM needs are better understood by Senior-level managers as they form the decision making team, which enables the firm to be competitive. They should show their commitment, by guiding and leading an effective SCRM implementation process (Dath et al., 2009; KPMG, 2010). Mitigation Strategies and Risk Management Process1 (MS-RMP) In the competitive and unstable environment, every firm inevitably faces risk. The risk in the organization cannot be avoided, but it can be curtailed to a certain degree. The organization can identify deviation, and implement appropriate strategies and convert negative events into positive or near positive outcomes (Sodhi and Chopra, 2004; Oke and Gopalakrishnan, 2009; Tang and Nurmaya Musa, 2011; Radke and Tseng, 2012; Arcelus et al., 2012; Chen et al., 2013; Sawik, 2013; Kırılmaz and Erol, 2017; Tarei et al., 2020). In this study, the following sequence is followed for alleviating risk: • Risk Planning (RP) – establishment of plan of actions such as a business continuity plan. • Risk Monitoring (RM) – continuous assessment of control parameters, active involvement of risk managers. • Risk Avoidance (RA) – collaboration among trading partners, implementation of forecasting and advance planning systems. • Risk Sharing (RS) – equitable sharing of the impact of risk among the trading partners as per agreeable established policies, insurance, and fund provisioning to meet exigencies. Performance measures2 (PERF) The effectiveness and efficiency of any system can be estimated only when they are measured. It helps in benchmarking the firms’ capabilities with the competing firms. The SCRM implementation warrants inclusion of measures of performance from an internal perspective include the inter-functional, customer and partner perspectives. From the perspective of the manufacturer, the measures of performance are finance, customer perspective, 2 Indicates the factors adapted appropriately and suitably from the work of Dath et al. (2009). The considered items are suitably modified and additional items added to every factor. SCRM in Indian Manufacturing Industries 111 trading partner perspective, internal business perspective and innovation, and learning perspective (Dath et al., 2009). Conceptual Framework The factors identified in Sect. 2.1 are portrayed through the conceptual model, as shown in Fig. 1, from the perspectives of the manufacturers (including OEMs and Suppliers) (Shenoi et al., 2016). However, they do not provide a comprehensive questionnaire based on the conceptual framework. They hypothesized the framework. In the model depicted below is a mapping from the SCRM constructs (the Independent Variables (IVs)) and the performance measures, i.e., the Dependent Variable(DV). The items in the middle are the mediator(s). The mediators cause the relationship between the IV and DV to change the pathway depending on the mediating variable’s participation in the relationship. In this study, it is evident that the presence of the mediators (namely, risk planning, risk monitoring, risk avoidance, and risk sharing described in Sect. 2.1) has a mediating effect on the Fig. 1 The framework of SCRM with respect to Manufacturers (Shenoi et al., 2016) 112 V. Viswanath Shenoi et al. association between the IVs and the DV taken one at a time (refer Fig. 1). Similar to SCM, the SCRM also involves managing complex flow of information and understanding the irrelevance. The SCRM gained importance due to the following (Faisal et al., 2007): • Focus on core-competencies. • Elimination of geographical boundaries for establishing partnerships. • Escalating disruptions in the supply chain due to man-made attacks and due to natural disasters. • Reduction in the supplier base attributed to JIT production. • Turbulent nature of the economy due to strong linkage to commodities. Lavastre et al. (2012) identified techniques for minimizing supply chain risks by considering the attitudes of the managers towards risk and the tools used to understand the risk. 2.2 Fuzzy Cognitive Map (FCM) and Its Applications in the Present Study: An Overview The findings from the empirical study help us to develop the Fuzzy Cognitive Map (FCM) needed to predict risks and suggest mitigation strategies. In the Sect. 2.1 and section “Description of the Construct of SCRM: Manufacturer’s Perspective”, we have identified the factors that affect the Supply chain. In the globalized environment, it is necessary to identify early warning signals and implement appropriate strategic decisions to meet the exigencies. The objective is to develop a fuzzy model to identify and predict all the plausible risks based on the instantaneous risk vector. We utilize the responses of the empirical study for the construction of FCM. The factors of SCRM represent the states of the FCM. FCM is used to represent the overall behavior of the dynamic supply chain system. The instantaneous risk vector is passed on to the dynamic system to identify all plausible risks until saturation. Risk Modeling: Graph Theory Approach The research contributions to managing supply chain risks, to date, are based on the identification of sources of risk, and recommendation of corresponding mitigation strategies. However, Sodhi and Chopra (2004) proposed that risks are dependent on each other, and therefore, it is necessary to understand the interdependencies between the risk constructs. Graph modeling seems to be an appropriate method for understanding the interdependencies. Graph modeling approaches have been applied to coordination, customer sensitivity, risk alleviation, and risk mitigation (Kaur et al., 2006; Faisal et al., 2007). The other cognitive approach widely used is neural regression. SCRM in Indian Manufacturing Industries 113 Using fuzzy logic and Analytic Hierarchical Process (AHP), Kaur et al. (2007) evaluated the extent of coordination among the trading partners in the supply chain. An integrated approach using FCM and the fuzzy soft set model is proposed to solve the supplier selection problem. The approach utilizes the feedback effect among the criteria and considers the uncertainties in the decision-making process. Micheli et al. (2014) developed an analytical model using stochastic integer linear programming approach that incorporated supply chain managers, judgments using utility functions, and the model employed fuzzy-extended pairwise comparisons for choosing the mitigation strategy. However, the application of FCM in SCRM context is sparsely available. In our study, FCM is utilized to construct the dynamical system model of risk present in the supply chain. Recurrent structures such as FCM help us to establish cause-effect relationships, as the system variables are identified to the graph nodes. The proposed model helps to determine all the plausible risks which may appear in a supply chain and suggests strategies to mitigate risk through a fuzzy approach. Basics of Fuzzy Cognitive Maps The Fuzzy Cognitive Map concept, extended by Kosko (1986), used fuzzy causal functions considering numbers in the interval [−1, 1]. Its application is extended to decision making (Stylios et al., 2008), performance measurement (Glykas, 2013), supplier selection (Shaw et al., 2012), and modeling of resilient supply chain network design (Kristianto et al., 2014). FCM provides a quick solution while modeling complex systems with lesser computational effort. The relationships between variables obtained constitutes the matrix of edges known as an adjacency model, representing the overall behavior of the system. A non-linear activation function is used for transformation of the path and is directed towards them into a value [0, 1] or [−1, 0, 1]. Cognitive mapping of FCMs helps the respondent to have an awareness of their prediction model by drawing, capturing and transferring causal knowledge. It enables a way to elicit, capture and transfer causal knowledge. Maps are constructed in various ways, such as conducting interviews, analyzing texts, capturing participant perceptions through a questionnaire, and discussing in groups, to name a few. Perceptions from varied and diverse sources are integrated to deal with limitations of expert opinions. FCMs also allow quantitative analysis and interpretation of quasidynamic behavior and support in making timely decisions. While adopting FCM, modelers can use all combinations of input variables or suggest alternative system descriptions while addressing complex problems or linking internally consistent future state. 114 V. Viswanath Shenoi et al. A1 e14 A4 e31 R1 e43 e12 A2 e23 R2 A3 Fig. 2 Risk interdependency graph The Proposed Model: Mapping Supply Chain Risks and Mitigation Strategies In the present study, a holistic approach is adapted to understand the interdependency of risks and associate them with a mitigation strategy via the concept of Fuzzy Cognitive Maps. Bivariate correlation on the responses helps us to understand the interdependency of the risk constructs on performance. The relationship is established for those correlations of the constructs which are significant. The impact of risk on the supply chain is categorized into two, namely, disruption and lower performance. For example, Ai indicates a risk construct in the supply chain, R1 indicates supply chain disruption and R2 indicates lower supply chain performance. The graph formulation is shown in Fig. 2. The Fuzzy construction of risk interdependency graphs is dealt in Sect. 2.2. In Fig. 2, Ai → Aj denotes that the risk construct Ai influences the risk construct Aj . Ai → R1 indicates that the risk construct Ai leads to supply chain disruption. Similarly, Ai → R2 indicates that the risk Ai leads to lower supply chain performance. Similarly, the association between the risks, mitigation strategies, supply chain disruption, and supply chain performance can be obtained. The graph formulation is shown in Fig. 3. In Fig. 3, Mi → Aj denotes the mitigation strategy, Mi mitigates the risk construct Aj , abates supply chain disruption (R1 ) and enhances supply chain performance (R2 ). Construction of FCM “FCM is a directed graph with nodes referred to as policies, events, etc., and causalities as edges. It represents the causal relationship between concepts” (Kandasamy et al., 2007). From the Bivariate Correlation established in Tables 15 and 16, and the rules defined for establishing the relationships, the FCM is developed. The directed edge eij from causal concept Ai to concept Aj measures the extent of influence SCRM in Indian Manufacturing Industries 115 R2 M1 R1 A1 A3 M2 A2 A4 Fig. 3 Risk-mitigation interdependency graph of Ai on Aj . The edge eij takes values in the fuzzy interval [0,1]. If eij = 0 indicates no causality, whereas eij > 0 indicates causal influence. The influence of concept Aj increases as Ai increases. if eij < 0, indicates causal decrease or negative causality i.e., the influence of concept Aj decreases as Ai increases. Consider the nodes A1 , A2 , . . . , An of the FCM. The edge eij represents the causal relationship between causal concept Ai to concept Aj , its value is one when there is significant correlation, otherwise it is zero. The adjacency matrix of FCM has been constructed by observing Table 16 in respect of the presence/absence of significant correlation; eij = ⎧ ⎨0 ⎩ 1 no significant correlation exists between causal concept Ai and concept Aj ; significant correlation exists between causal concept Ai and concept Aj . (1) A symmetric Matrix E is obtained (see Table 3) and the entries in the main diagonal are equal to 1. The relation represented is reflexive and symmetric. Hence, it is a compatibility relation (Kandasamy et al., 2007). The diagonal entries are zero, as they are self-contained. eij = 0 if i = j To identify and understand the causal pathways and to initiate an appropriate measure to reduce the impact of disruption, early detection of the risk is essential. Appropriate mitigation strategies need to be suggested and adopted to reduce the impact of risk on the supply chain. The risk factors, namely supply chain disruption and supply chain performance, augmented with the concept of risk factors in the adjacency matrix. The dimension of adjacency matrix E is n × n. The 116 V. Viswanath Shenoi et al. relationship between the concept Ai and the risk factor-disruption is captured in r1 : (Ai , r1 ) = 1 0 ifAi disrupts the supply chain; ifAi does not disrupt the supply chain. (2) The relationship between the concept Ai and its risk factor-performance is captured in r2 : (Ai , r2 ) = 1 0 if Ai impacts the performance of the supply chain; if Ai does not impact the performance of the supply chain. (3) Thus, the dimension of adjacency matrix changes from n × n to (n + 2) × (n + 2). The new augmented matrix is denoted by E . The mapping between risk items and mitigation strategies needs to be established for studying the mitigating effects. Let B = (b1 , b2 , . . . , bm ) be the set of mitigation strategies. The relationship matrix M is constructed as follows: ⎧ ⎨1 if bi reduces the impact of risk due to Ai and; enhances performance (Ai , bi ) = (4) ⎩ if b does not reduce the impact of risk due to A 0 i i . Thus the dimension of the matrix is n × m. To identify the appropriate mitigation strategy, it is necessary to establish the relationship between the impact factor and mitigation strategy: (r1 , bi ) = n −1 if bi reduces disruption due to Ai and enhances ; performance (5) −1 if bi reduces risk due to Ai and enhances perfor; mance (6) i=1 (r2 , bi ) = n i=1 After augmenting the impact factors with the mitigation strategies, the dimension of new relationship matrix M is (n + 2) × m. Prediction of Future Risks Based on the Current State of Risk Observance Let A1 , A2 , . . . , An be the nodes of an FCM. A = (a1 , a2 , . . . , an ), where ai ∈ {0,1}. A is called an instantaneous state vector and it denotes the on-off position of SCRM in Indian Manufacturing Industries the node in an instant of time: 0 if ai off, no risk is perceived for the concept Ai ; ai = 1 if ai on, risk is perceived for the concept Ai . 117 (7) When A = (a1 , a2 , . . . , an+2 ) vector is passed into a dynamical system of matrix E . The resultant vector A ∗ E = (a1 , a2 , . . . , an+2 ) is subject to threshold of 1. The vectors are updated and renamed as D = (d1 , d2 , . . . , dn+2 ). We denote the above operation by (a1 , a2 , . . . , an+2 ) (b1 , b2 , . . . , bn+2 ). The symbol represents the resultant vector has been subject to a threshold of 1 and updated. The above process is repeated till the FCM state vector reaches the equilibrium state of a dynamical system, i.e., when there is no change in the resultant state vector’s component-wise from that of the immediate preceding resultant vector. Then the resultant vector obtained is called the fixed point. The fixed point is the saturation point of risk that may lead to a total breakdown of the supply chain. Identification and Effectiveness of Mitigation Strategy for Risk During the Run To identify the appropriate mitigation strategy, the vector D = (d1 , d2 , . . . , dn ) is multiplied with matrix M. Thus the resultant vector, B = (b1 , b2 , . . . , bn ), is obtained. The set of bi ’s, which are equal to 1 or non-zero values, are the mitigation strategies to be implemented at that level to alleviate the impact of risk and enhance the supply chain performance. To study the effectiveness of the mitigation strategy B , we need to find the product of the vector, D = (d1 , d2 , . . . , dn+2 ), and the matrix M. If the resultant vector is zero or a negative value, then it signifies that all the risks are alleviated. 2.3 Fuzzy TOPSIS Approach In order to compare the results of the FCM approach, we consider the Fuzzy TOPSIS approach. Identification of risks present in the supply chain and appropriate mitigation strategies to curtail them is a Multi-Criteria Decision Making (MCDM) problem. A solution to the MCDM problem can be achieved through TOPSIS (Technique for Order Performance by Similarity to Ideal Solution) approach (Rostamzadeh et al., 2018; Al Zubayer et al., 2019). This technique suggests the best solution alternatives closest to the Positive Ideal Solution (PIS) and distant from the Negative Ideal Solution (NIS). The benefit criteria is maximized and cost criteria is minimized by PIS and NIS. This section deals with TOPSIS applied to the fuzzy environment as proposed by Shemshadi et al. (2011) and Zadeh (1965) to map the linguistic variables to the numerical variables. 118 V. Viswanath Shenoi et al. Definition 1 Let a = (a1 , a2 , a3 ) and b = (b1 , b2 , b3 ) be two fuzzy triangular numbers, then the distance between them is determined by the vertex method given in equation 8 (Dağdeviren et al., 2009). d(a, b) = 1 [(b1 − a1 )2 + (b2 − a2 )2 + (b3 − a3 )2 ] 3 (8) Definition 2 The importance values for each criterion, the weighted normalised fuzzy decision matrix is constructed as: V = [vij ]m×n i = 1, 2, . . . , n j = 1, 2, . . . , m vij = rij ⊗ wi (9) The steps of Fuzzy TOPSIS is as follows: Step 1: The linguistic ratings (xij i = 1, 2, . . . , n j = 1, 2, . . . , m) for the alternatives with respect to criteria R is obtained. To normalise the decision matrix R, let: xij = (aij , bij , cij ) xj = (aj , bj , cj ) xj∗ = (aj∗ , bj∗ , cj∗ ) then rij = Step 2: Step 3: Step 4: ⎧ ⎪ xij ⎪ ⎪ ⎪ ⎪ ⎨ x∗ j ⎪ ⎪ xj ⎪ ⎪ ⎪ ⎩ xij = aij bij cij , , aj∗ bj∗ cj∗ = a j b j cj , , aij bij cij ; (10) The weighted normalized fuzzy decision matrix is calculated. The weighted normalized value vij is calculated by using the equation 9. The Fuzzy Positive Ideal Solution (FPIS) and the Fuzzy Negative Ideal Solution (FNIS) is identified. As this work is based on the risk criteria, FPIS will have maximum value and FNIS will have minimum value. The distance is calculated for each of the alternative from FPIS and FNIS as mentioned in equations 11 and 12. Dj∗ = m (vij , vj∗ ) j = 1, 2, . . . , m (11) j = 1, 2, . . . , m (12) j =1 Dj− = m (vij , vj− ) j =1 SCRM in Indian Manufacturing Industries Step 5: 119 The similarity to ideal solution is calculated. CCj = Step 6: Dj− Dj− + Dj∗ (13) The alternatives are ranked according to CCj in decreasing order. 2.4 Research Gaps, Hypotheses and Contributions Literature to date concerning India is majorly on supply chain flexibility (Khan and Pillania, 2008), and alignment of supply chain strategy with the business strategy (Dath et al., 2010). However, articles based on SCRM practices in Indian scenario are rare. Hence, the present study provides insight into SCRM research in the Indian context, considering the vibrant expansion of industries in India. Researchers observed that SCRM is multifarious, involves numerous interrelated risk factors, and needs a rigorous study to address it. Risk mitigation strategies differ based on the perception of risk and the implementation strategy that could extend for a short, medium or long term. Based on the literature review, 12 effective techniques widely employed by mangers for minimizing supply chain risks have been identified. The techniques have been grouped into four categories, namely risk planning, risk monitoring, risk avoidance, and risk sharing. The above discussion strengthens the argument that the four mitigation strategy constructs have a substantial effect on the relationship between the eight SCRM constructs and the performance measures. The above observations lead us to 32 hypotheses from the perspective of manufacturers. For example, H1a – corresponds to the effect of the mitigation strategy, namely risk planning on the risk construct ‘Demand side risk’. Similarly, other hypotheses can be developed based on mitigation strategies with the risk constructs taken one at a time. Table 1 summarizes the hypotheses conceived in this work. The hypotheses conceived shows the effect of each of the individual mediators (namely, risk planning, risk monitoring, risk avoidance, and risk sharing) on the relationship between Independent variables (IVs), taken one at a time, and Dependent Variable (DV), namely, performance measures. Thus the hypotheses H1a, H1b, H1c, H1d, H1e, H1f, H1g, and H1h correspond to the effect of mediator variable, Risk avoidance (RA), on Demand Side Risk, Supply Side Risk, Logistic Risk, Regulatory, Legal and Bureaucratic Risk, Infrastructure Risk, Stock/Data Management risk, Environmental Risk, and Financial Risk respectively. The hypotheses H2, H3, and H4 correspond to the effect of mediating variables, Risk Planning, Risk Monitoring, Risk Avoidance and Risk Sharing, on the individual risk constructs. Further, the FCM is developed by using the bivariate correlation among the risk variables. The study also illustrates practitioners and researchers applying concepts such as FCM and Fuzzy TOPSIS for the identification and ranking of strategies. 120 V. Viswanath Shenoi et al. Table 1 The list of formulated hypotheses: the perspective of manufacturers Mediator Risk Planning (RP): (corresponding hypothesis are H1a, H1b, H1c, H1d, H1e, H1f, H1g, H1h) Hypothesis RP has a significant and mediating effect on the relationship between the individual independent variables taken one at a time and the measures of performance. Refer to Fig. 1. Risk Monitoring (RM): (corresponding hypothesis are H2a, H2b, H2c, H2d, H2e, H2f, H2g, H2h) RM has a significant and mediating effect on the relationship between the individual independent variables taken one at a time and the measures of performance. Refer to Fig. 1. Risk Avoidance (RA): (corresponding hypothesis are H3a, H3b, H3c, H3d, H3e, H3f, H3g, H3h) RA has a significant and mediating effect on the relationship between the individual independent variables taken one at a time and the measures of performance. Refer to Fig. 1. Risk Sharing (RS): (corresponding hypothesis are H4a, H4b, H4c, H4d, H4e, H4f, H4g, H4h) RS has a significant and mediating effect on the relationship between the individual independent variables taken one at a time and the measures of performance. Refer to Fig. 1. 3 Methods The framework of the study is presented in Fig. 4. 3.1 Data Description This study helps practitioners understand the likelihood of risks and their severity in the practitioners’ respective companies. The responses to the questionnaire3 from the practicing managers serves as the data for our study. 3.2 Survey Description The population considered for the study includes OEMs and suppliers who are the manufacturers in India. The final product manufacturers are the OEMs in various sectors, namely, the automotive (e.g., manufacturers of two and fourwheeler), the heavy engineering (e.g., manufactures of heavy machinery), home appliances (e.g., manufacturers of washing machines), and general engineering. The respondents considered for our study includes Vice-Presidents and Senior personnel in the purchasing and materials management departments. Numerous respondents were contacted, and the questionnaire was shared with around 120 shortlisted respondents. Around 40 respondents are from OEMs, and 80 respondents 3 https://tinyurl.com/scrmquestionnaire SCRM in Indian Manufacturing Industries 121 Fig. 4 Framework of the study are suppliers to OEMs. The contacted respondents work in various industrial sectors mentioned in the sample profile summarized in Table 2 below. Similarly, for the supplier perspective studies, sub-assembly suppliers to OEMs in the above-mentioned sectors were shortlisted. The supplier respondents are manufacturers of fasteners, brake linings, wheels, electrical components such as coils, motors, etc. In the supplier perspective study, out of 80 suppliers contacted, 52 suppliers responded. A majority of the respondents (44) are from the organi- 122 V. Viswanath Shenoi et al. Table 2 OEM respondents’ categorization Company categorization Automobile (two and four-wheeler manufacturers) Heavy vehicles (trucks and medium sized vehicles manufacturers) Heavy engineering (earth moving equipment and heavy machinery manufacturers) General engineering (automobile components and spare part manufacturers) Number of respondents OEM Supplier 7 12 6 10 2 10 18 20 zations that supply sub-assemblies and parts to OEM’s of two and four-wheeler vehicles, large capacity vehicles and earth-moving equipment manufacturers. The questionnaires were circulated to the respondents in three different forms, namely, direct interview, E-mail and web forms. 3.3 Scales Used to Measure the Latent Variables The supply chain managers underestimate risk in terms of their likelihood and impact (Thun and Hoenig, 2011). The importance of development of instruments is emphasized by the researchers for theory building (Dath et al., 2010). The survey instruments help the managers to identify the aspects of SCRM and empirically validate the SCRM constructs from the perspective of the manufacturers. The exhaustive review of the literature resulted in the development of the instrument. An instrument with 1 indicating ‘very low’ and 5 indicating ‘very high’ (referred to as five point Likert scale) has been adopted to study the likelihood of the risk and its severity in this work. 3.4 Construction of FCM for SCRM As discussed in Sect. 2.2, the adjacency matrices are computed for the risk constructs. The correlation matrix is given in Table 16. In this study, Ai ’s are considered as risk constructs, namely, QDSR, QSSR, QLR, QRLB, QIR, QSDM, QER, and QFR. ‘Q’ is prefixed with the constructs label to indicate the severity of the impact due to that risk. The entries in Table 3 corresponds to the significant correlations between the variables as observed from Table 16 along with equation 1, 2 and 3. ‘Q’ is prefixed with the mitigation strategies to indicate the impact of implementing the alleviation strategy. Mi ’s are considered as QRP, QRM, QRA, and QRS. The relationship matrix presented in Table 4 is also generated from Table 16, along with SCRM in Indian Manufacturing Industries 123 Table 3 Adjacency matrix E : FCM of concepts of risk constructs ∗ QDSR QSSR QLR QRLB QIR QSDM QER QFR R1 R2 QDSR 0 1 1 1 1 1 1 0 1 1 QSSR 1 0 1 1 1 1 1 1 1 0 QLR 1 1 0 1 1 1 1 1 1 0 Table 4 Relationship matrix M : concepts of risk vs mitigation strategies QRLB 1 1 1 0 1 1 1 1 0 1 QIR 1 1 1 1 0 1 1 1 0 1 QSDM 1 1 1 1 1 0 1 1 1 0 ∗ QDSR QSSR QLR QRLB QIR QSDM QER QFR R1 R2 QER 1 1 1 1 1 1 0 1 1 0 QRP 1 1 0 0 1 1 1 0 −3 −2 QFR 0 1 1 1 1 1 1 0 1 1 QRM 1 1 1 0 1 1 1 1 −3 −4 R1 1 1 1 0 0 1 1 1 0 1 QRA 1 1 1 1 0 1 1 0 −3 −3 R2 1 0 0 1 1 0 0 1 1 0 QRS 1 1 1 0 0 1 1 1 −3 −3 equations 4, 5 and 6. Figure 5 represents the construction of the FCM model of risk constructs. Figure 6 represents the relationship between risk constructs, mitigation strategies, and the nature of the impact on the supply chain. 3.5 Statistical Analysis Empirical Validation of the Proposed SCRM Constructs After data collection, processes including refinement, modification, and finalization of the measurements are undertaken. The data is subjected to factor analysis in order to discover the latent variables with the help of item-factor loadings. The questionnaires are subjected to reliability and validity tests, for ensuring standardization and making them operational (Sureshchandar et al., 2001; Dath et al., 2010). Confirmatory Factor Analysis (CFA) method is followed in the present study. The CFA approach helps the researchers confirm the hypotheses and test the established relationship between observed and latent variables as a specified model a priori by the researcher. Since the researchers possesses the knowledge of the 124 V. Viswanath Shenoi et al. QDSR QSSR QFR QLR R1 QRLB R2 QIR QER QSDM Fig. 5 Risk interdependency graph observed variables, reliable indicators for each of the constructs, the CFA technique has been adopted. Reliability and Validity Tests On freezing the scale of measurement, the researchers need to ensure the construct validity of the instrument and gain confidence that the inferences derived based on the proposed model reflects reality. The fundamental prerequisite for construct validity is the unidimensionality of the measure. CFA is done for all the constructs. A value of 0.9 or higher for the Comparative Fit Index (CFI) suggests that the model framework has an adequate fit, and unidimensionality exists (Sureshchandar et al., 2001). The Root Mean Error of Approximation (RMSEA) is an additional measure for model fit adequacy. The maximum value for RMSEA deemed to be acceptable is 0.1 (Hair et al., 2006). The consistency of the instrument is a measure of Reliability, also referred to as internal consistency to measure the intended issue. The measure of internal SCRM in Indian Manufacturing Industries 125 Ai QDSR QSSR Mi QRP QLR QRM R1 QRLB QRA R2 QIR QRS QSDM QER QFR Fig. 6 Risk-mitigation interdependency graph consistency is Cronbach’s alpha (α) (Cronbach, 1951), and a value of 0.7 and above is acceptable. Face and content validity indicates the extent to which the research instrument depicts the concepts (Sureshchandar et al., 2001) and determines the face value of a good representation of the construct (Kaplan and Saccuzzo, 2012). Review of the questionnaire by expert professors and industry personnel ensures face validity. The 126 V. Viswanath Shenoi et al. extent to which the factors converge is indicated by convergent validity and a value of 0.9 and above for Bentler Bonet Fit Index (BBI) satisfies the requirement. In the present study, the instrument is developed for manufacturers, which include suppliers and OEMs. The items contained in instruments are drawn from the literature. 4 Results 4.1 Factor Loadings Table 5 presents the Factor loadings containing the values of various indices such as BBI, CFI, Chronbach’s alpha and RMSEA. The values are above the limits satisfying the required validity and reliability requirements. Top Management Commitment (TMC) and Mitigation strategies (MS) have RMSEA values above 0.1, but the values of other indices such as BBI and CFI are in range. Tables 5, 15 and 16 indicate the presence of discriminant validity (Crocker and Algina, 1986). 4.2 Sobel Test for Mediation Sobel test is performed to test whether a mediator realizes the influence of an IV to a DV (Preacher and Hayes, 2008). The parameters in the Sobel test are summarized in the Table 6 below. On performing the Sobel test (Preacher, 2010), Test Statistic (TS), and Standard Error (SE) with associated p-value is obtained. If the p-value falls below the established alpha level of 0.05, indicating the association between the IV and DV (in this case, risk constructs, influence on performance), the impact of the risk construct is significantly reduced by the inclusion of the mediator (in this case mitigation strategies). It is a evidence of mediation. In the tables below, R is referred to risk, MV is referred to mediating variable, and P is the performance. The stepwise regression method is implemented as follows (Tables 7, 8, 9, 10, 11, 12, 13, 14): (a) Risk(R) is considered as an independent variable, and the mediating variable(MV) is considered as a dependent variable and is denoted by R→MV(a). (b) Risk(R) and mediating variable(MV) are considered as independent variables, and the performance measures(P) is considered as a dependent variable, and is denoted by R& MV → P(b). 4.3 Bivariate Correlation Between the Constructs The Bivariate Correlation is performed for understanding the relationship between the likelihood of the risk and performance. Also, the Bivariate Correlation is SCRM in Indian Manufacturing Industries 127 Table 5 The Construct of SCRM: item-factor loadings and fit indices Item number Item loadings BBI CFI α RMSEA 1 0.45 0.97 0.99 0.73 0.08 2 3 4 1 2 3 4 1 2 3 4 5 1 0.88 0.94 0.56 0.93 1.06 0.93 0.80 0.73 0.75 0.77 0.93 1.04 0.72 0.98 0.99 0.90 0.11 0.96 0.97 0.90 0.14 0.87 0.89 0.77 0.00 2 0.88 Infrastructure Risk(IR) 3 4 1 2 3 0.53 0.51 0.71 0.73 0.98 0.99 1.00 0.82 0.00 Stock and Data Management (SDM) 1 0.92 0.99 1.00 0.86 0.00 2 3 0.83 0.71 1 1.03 0.98 0.99 0.898 0.16 2 3 4 1 2 3 1 2 3 4 5 6 7 8 9 1.00 1.04 0.88 0.61 0.66 0.85 0.79 0.88 0.57 0.77 0.80 0.76 0.79 0.80 0.72 1.00 1.00 0.84 0.00 0.94 0.96 0.94 0.13 Constructs Demand Side (DSR) Risk Supply Side Risk (SSR) Logistic Risk (LR) Regulatory, Legal and Bureaucratic Risk (RLB) Environmental (ER) Risk Financial Risk (FR) Top Management Commitment (TMC) (continued) 128 V. Viswanath Shenoi et al. Table 5 continued. Item Item Constructs number loadings BBI Mitigation Strategies and Risk Management Process (MS-RMP) Risk Planning (RP) 1 0.92 1.00 2 0.61 3 0.69 4 0.64 Risk Monitoring (RM) 1 0.88 1.00 2 0.64 3 0.63 Risk Avoidance (RA) 1 0.81 1.00 2 0.73 3 1.00 Risk Sharing (RS) 1 0.84 1.00 2 0.86 3 0.80 1 0.61 0.90 Financial Performance 2 0.67 3 0.89 4 0.56 5 0.54 6 0.44 Customer perspective 1 0.70 0.97 performance 2 0.73 3 0.79 4 0.79 5 0.86 6 0.54 Trading partner perfor1 0.79 0.95 mance 2 0.89 3 1.00 4 0.83 5 0.87 Internal business 1 0.68 0.98 perspective performance 2 0.85 3 0.75 4 0.77 5 0.76 6 0.79 Innovation and learning 1 0.77 1.00 perspective performance 2 0.69 3 0.86 CFI α RMSEA 1.00 0.783 0.00 1.00 0.74 0.00 1.00 0.86 0.00 1.00 0.87 0.00 0.91 0.83 0.12 0.98 0.93 0.11 0.96 0.95 0.09 1.00 0.92 0.07 1.00 0.93 0.00 Notes: Acceptable limits: CFA Index: BBI > 0.9; CFI > 0.9; α > 0.6; RMSEA< 0.1 SCRM in Indian Manufacturing Industries 129 Table 6 Parameters of Sobel Test Parameters R MV R → MV (a) Sa R& MV → P(b) Sb TS SE Details Risk construct Mediating Variable Raw (unstandardized) regression coefficient for the association between Risk construct-R and Mediating Variable (MV) Standard error of a Raw (unstandardized) regression coefficient for the association between the Risk-R and Mediating Variable (MV) and the Dependent Variable Performance (P) Standard error of b Test statistic Standard error of Sobel test Table 7 The effect of mediating variables on Demand side risk (DSR) Risk-R DSR DSR DSR DSR MV RP RM RA RS R → MV(a) 3.87 4.09 4.21 4.50 Sa 0.23 0.27 0.30 0.29 R&MV → P(b) 0.44 0.49 0.36 0.40 SOBEL TEST TS SE 4.69 0.36 6.35 0.31 4.82 0.31 5.36 0.33 p-value 0.00 0.00 0.00 0.00 Sb 0.09 0.07 0.07 0.07 SOBEL TEST TS SE 4.74 0.34 6.40 0.29 4.23 0.32 5.32 0.31 p-value 0.00 0.00 0.00 0.00 Sb 0.10 0.07 0.08 0.07 SOBEL TEST TS SE 4.31 0.41 6.46 0.29 4.35 0.34 5.60 0.32 p-value 0.00 0.00 0.00 0.00 Sb 0.09 0.07 0.07 0.07 Table 8 The effect of mediating variables on Supply Side Risk(SSR) Risk-R SSR SSR SSR SSR MV RP RM RA RS R → MV(a) 3.74 3.90 3.90 4.35 Sa 0.19 0.23 0.25 0.24 R&MV → P(b) 0.44 0.49 0.36 0.39 Table 9 The effect of mediating variables on Logistic Risk Risk-R LR LR LR LR MV RP RM RA RS R → MV(a) 4.08 3.89 4.14 4.47 Sa 0.18 0.23 0.24 0.23 R&MV → P(b) 0.44 0.49 0.36 0.41 130 V. Viswanath Shenoi et al. Table 10 The effect of mediating variables on Regulatory, Legal and Bureaucratic risk(RLB) Risk-R RLB RLB RLB RLB MV RP RM RA RS R → MV(a) 4.08 3.89 4.14 4.47 Sa 0.18 0.23 0.24 0.23 R&MV → P(b) 0.44 0.49 0.36 0.41 Sb 0.11 0.07 0.08 0.08 SOBEL TEST TS SE 3.93 0.45 6.46 0.29 4.35 0.34 4.95 0.36 p-value 0.00 0.00 0.00 0.00 SOBEL TEST TS SE 4.94 0.34 6.66 0.28 5.11 0.29 5.64 0.31 p-value 0.00 0.00 0.00 0.00 Table 11 The effect of mediating variables on Infrastructure Risk(IR) Risk-R IR IR IR IR MV RP RM RA RS R → MV(a) 3.76 3.78 3.99 4.19 Sa 0.19 0.23 0.26 0.25 R&MV → P(b) 0.46 0.51 0.38 0.42 Sb 0.09 0.07 0.07 0.07 Table 12 The effect of mediating variables on Stock/Data Management Risk (SDM) Risk-R SDM SDM SDM SDM MV RP RM RA RS R → MV(a) 3.94 3.84 4.35 4.37 Sa 0.20 0.25 0.26 0.26 R&MV → P(b) 0.44 0.49 0.36 0.41 Sb 0.10 0.07 0.08 0.07 SOBEL TEST TS SE 4.29 0.40 6.36 0.29 4.34 0.36 5.53 0.32 p-value 0.00 0.00 0.00 0.00 Table 13 The effect of mediating variables on Environmental Risk (ER) Risk-R ER ER ER ER MV RP RM RA RS R → MV(a) 3.77 3.72 3.85 4.14 Sa 0.14 0.18 0.19 0.19 R&MV → P(b) 0.43 0.49 0.35 0.40 Sb 0.10 0.07 0.07 0.07 SOBEL TEST TS SE 4.24 0.38 6.63 0.27 4.85 0.27 5.52 0.29 p-value 0.00 0.00 0.00 0.00 SOBEL TEST TS SE 4.93 0.33 6.35 0.29 4.96 0.27 5.39 0.30 p-value 0.00 0.00 0.00 0.00 Table 14 The effect of mediating variables on Financial Risk(FR) Risk-R FR FR FR FR MV RP RM RA RS R → MV(a) 3.60 3.90 3.75 4.08 Sa 0.19 0.23 0.26 0.25 R&MV → P(b) 0.46 0.48 0.37 0.40 Sb 0.09 0.07 0.07 0.07 performed for risk construct and performance. The Bivariate correlation helps the researchers to understand the inter-dependency between the variables (Tables 15 and 16). IDSR 1 0.61a 0.68a 0.31a 0.47a 0.29a 0.49a 0.11 −0.31a −0.15 −0.29a −0.29a −0.42a ILR 1 0.35a 0.65a 0.15 0.69a 0.29a −0.36a −0.21 −0.27b −0.44a −0.38a ISSR 1 0.60a 0.47a 0.56a 0.09 0.35a 0.26b −0.12 −0.16 −0.09 −0.27b −0.30a 1 0.71a −0.04 0.33a 0.15 −0.15 −0.21 −0.16 −0.16 −0.33a IRLB 1 0.09 0.56a 0.06 −0.07 −0.07 −0.02 −0.11 −0.27b IIR 1 0.16 −0.31a 0.04 0.25b 0.09 0.12 0.21 ISDM IER 1 0.09 −0.33a −0.20 −0.19 −0.39a −0.46a ‘I’ is prefixed with the constructs label to indicate the measure of likelihood a Correlation is significant at 0.01 level (2-tailed) b Correlation is significant at 0.05 level (2-tailed) Note: * IDSR ISSR ILR IRLB IIR ISDM IER IFR IRP IRM IRA IRS PERF 1 −0.14 −0.21 −0.13 −0.32a −0.46a IFR Table 15 Bi-Variate Correlation between the likelihood occurrence of risk constructs and performance 1 0.71a 0.78a 0.79a 0.53a IRP 1 0.78a 0.76a 0.59a IRM 1 0.82a 0.64a IRA 1 0.68a IRS 1 PERF SCRM in Indian Manufacturing Industries 131 QDSR 1 0.62a 0.63a 0.30a 0.38a 0.42a 0.44a 0.27b −0.20 −0.22b −0.22b −0.28a −0.28a −0.31a QLR 1 0.62a 0.77a 0.71a 0.81a 0.42a −0.22b −0.41a −0.17 −0.31a −0.34a −0.29a QSSR 1 0.71a 0.65a 0.63a 0.61a 0.55a 0.34a −0.13 −0.20 −0.18 −0.22b −0.28a −0.34a 1 0.78a 0.58a 0.56a 0.39a −0.17 −0.23b −0.16 −0.28a −0.22b −0.35a QRLB 1 0.69a 0.72a 0.44a −0.09 −0.20 −0.12 −0.23b −0.20 −0.24b QIR 1 0.55a 0.35a −0.23b −0.29a −0.14 −0.36a −0.26b −0.30a QSDM a Correlation ‘Q’ is prefixed with the constructs label to indicate the severity of impact is significant at 0.01 level (2-tailed) b Correlation is significant at 0.05 level (2-tailed) Note: * QDSR QSSR QLR QRLB QIR QSDM QER QFR QTMC QRP QRM QRA QRS PERF 1 0.50a −0.25b −0.29a −0.12 −0.23b −0.25b −0.32a QER Table 16 Bi-variate Correlation with respect to risk constructs and performance 1 −0.38a −0.11 −0.19 −0.12 −0.15 −0.36a QFR 1 0.55a 0.67a 0.60a 0.65a 0.79a 1 0.76a 0.80a 0.77a 0.48a QTMC QRP 1 0.81a 0.77a 0.61a QRM 1 0.85a 0.51a QRA 1 0.56a QRS 1 PERF 132 V. Viswanath Shenoi et al. SCRM in Indian Manufacturing Industries 133 4.4 FCM Results Instantaneous vector Q1 represents the risk perceived during an instant of time. The instantaneous vector Q1 is passed on to the adjacency represented in Table 3. The resultant vector is threshold to one and is passed on to the adjacency matrix until equilibrium. Equilibrium is attained in the third iteration (refer to Table 17). The equilibrium vector obtained is passed on to the relationship matrix M . The results in Table 18 suggest that the risks are mitigated, and the impact of risks are alleviated to a greater extent, as the results show negative values or zeroes. The equilibrium vector obtained is passed on to the relationship matrix M. The results in Table 19 suggest the levels at which the individual mitigation strategy needs to be implemented to mitigate the risk and enhance the performance. The highest score of 7 is obtained for risk monitoring, suggesting that regular monitoring of the supply chain is essential to make proactive decisions. The next highest score of 6 is realized for risk avoidance and risk sharing, and a score of 5 for risk planning (Table 19). Table 17 Results of product of instantaneous vector Q1 and dynamical system matrix E Q1 RS1 Q2 =˜ RS1 RS2 Q3 =˜ RS2 RS3 QDSR 1 0 1 8 1 8 QSSR 0 1 1 7 1 8 QLR 0 1 1 7 1 8 QRLB 0 1 1 7 1 8 QIR 0 1 1 7 1 8 QSDM 0 1 1 7 1 8 QER 0 1 1 7 1 8 QFR 0 0 0 8 1 8 R1 0 1 1 6 1 7 R2 0 1 1 4 1 5 R1 1 R2 1 Table 18 Results of product of equilibrium vector Q3 and relationship matrix M RS3 QDSR 1 QRP 0 QSSR 1 QRM 0 QLR 1 QRA 0 QRLB 1 QRS 0 QIR 1 QSDM 1 QER 1 QFR 1 Table 19 Results of product of instantaneous vector Q1 and relationship matrix M RS3 QDSR 1 QRP 5 QSSR 1 QRM 7 QLR 1 QRA 6 QRLB 1 QRS 6 QIR 1 QSDM 1 QER 1 QFR 1 134 V. Viswanath Shenoi et al. 4.5 Fuzzy TOPSIS Approach The risk is categorized into four types the low likelihood and low impact, the low likelihood and high impact, the high likelihood and low impact, and high likelihood and high impact. In the proposed model, the normalized weights for each criterion is obtained from the likelihood of a risk. The ranking of the alternatives and the weights of the criteria are calculated based on Fuzzy TOPSIS approach. The theory of fuzzy sets reconciles the alternatives evaluation to deal with the uncertainty. The triangular fuzzy numbers used for the evaluation are represented in Fig. 7. In this problem, there are eight risk factors and four mitigation strategies as alternatives. The desire is to rank the alternatives using risk and to reduce the risk present in the supply chain. The steps involved in the ranking of the alternatives as follows: 1. There are four alternatives, namely risk planning, risk monitoring, risk avoidance and risk sharing. 2. Risk criteria considered are: • • • • • • • • Demand Side Risk (DSR) Supply Side Risk (SSR) Logistic Risk (LR) Regulatory, Legal and Bureaucratic risk (RLB) Infrastructure Risk (IR) Stock/Data Management risk (SDM) Environmental Risk (ER) Financial Risk (FR) 3. The weights of the criteria are calculated (see Table 20) 4. The alternatives are evaluated using the linguistic variables. The linguistic values are converted to appropriate equivalent fuzzy numbers (see Tables 21 and 22). 1.0 0.1 Very low Low 0.2 0.4 0.3 High 0.5 Fig. 7 Triangular fuzzy numbers (Liu et al., 2005) 0.6 Very high 0.7 0.8 0.9 SCRM in Indian Manufacturing Industries 135 Table 20 Total weights of the criteria (normalised) Criteria DSR SSR LR RLB IR SDM ER FR Weights 0.13 0.14 0.13 0.13 0.12 0.12 0.11 0.12 ER 4.24 6.63 4.85 5.52 FR 4.93 6.35 4.96 5.39 Table 21 Decision matrix for the alternatives RP RM RA RS DSR 4.69 6.35 4.82 5.36 SSR 4.74 6.4 4.23 5.32 Table 22 Relationship between linguistic parameters and fuzzy numbers LR 4.31 6.46 4.35 5.6 RLB 3.93 6.46 4.35 4.95 IR 4.94 6.66 5.11 5.64 Linguistic parameter Very Low (VL) Low (L) High (H) Very High (VH) SDM 4.29 6.36 4.34 5.53 Numerical value <4.69 <5.11 <6.35 >=6.35 Fuzzy number (0.1,0.2,0.3) (0.3,0.4,0.5) (0.5,0.6,0.7) (0.7,0.8,0.9) 5. The weighted decision matrix for alternatives are calculated (see Tables 23, 24, and 25). 6. FPIS and FNIS are calculated (see Table 26 for results) 7. The distance of each alternative is calculated from FPIS and FNIS to determine the closeness coefficient for each of the alternatives (see Table 27). 8. Rank of the alternatives with respect to their closeness coefficient is thus obtained and is shown in Table 28. 5 Discussion (Managerial and Theoretical Implications) 5.1 Validation of the Proposed Conceptual Framework From the manufacturer’s perspective, the effect of each mediator in the relationship between IVs is taken one at a time, and DV is individually studied through regression analysis using SPSS. Some of the items are eliminated during the validation process. The complete final instrument can be had from link.3 Tables 7, 8, 9, 10, 11, 12, 13, and 14 give the results of the effect of individual mediators on individual IVs taken one at a time. The Sobel test is done to analyze the RP RM RA RS 0.3 0.7 0.3 0.5 DSR 0.4 0.8 0.4 0.6 0.5 0.9 0.5 0.7 0.3 0.7 0.1 0.5 SSR 0.4 0.8 0.2 0.6 0.5 0.9 0.3 0.7 Table 23 Decision matrix for alternatives LR 0.1 0.7 0.1 0.5 0.2 0.8 0.2 0.6 0.3 0.9 0.3 0.7 RLB 0.1 0.7 0.1 0.3 0.2 0.8 0.2 0.4 0.3 0.9 0.3 0.5 IR 0.3 0.7 0.5 0.5 0.4 0.8 0.6 0.6 0.5 0.9 0.7 0.7 SDM 0.1 0.7 0.1 0.5 0.2 0.8 0.2 0.6 0.3 0.9 0.3 0.7 ER 0.1 0.7 0.3 0.5 0.2 0.8 0.4 0.6 0.3 0.9 0.5 0.7 FR 0.3 0.7 0.3 0.5 0.4 0.8 0.4 0.6 0.5 0.9 0.5 0.7 136 V. Viswanath Shenoi et al. RP RM RA RS DSR (0.13) 0.33 0.44 0.78 0.89 0.33 0.44 0.56 0.67 0.56 1 0.56 0.78 SSR (0.14) 0.33 0.44 0.78 0.89 0.11 0.22 0.56 0.67 0.56 1 0.33 0.78 LR (0.13) 0.11 0.22 0.78 0.89 0.11 0.22 0.56 0.67 0.33 1 0.33 0.78 Table 24 Normalised Fuzzy Decision matrix for alternatives RLB (0.13) 0.11 0.22 0.78 0.89 0.11 0.22 0.33 0.44 0.33 1 0.33 0.56 IR (0.12) 0.33 0.44 0.78 0.89 0.56 0.67 0.56 0.67 0.56 1 0.78 0.78 SDM (0.12) 0.11 0.22 0.33 0.78 0.89 1 0.11 0.22 0.33 0.56 0.67 0.78 ER (0.11) 0.11 0.22 0.78 0.89 0.33 0.44 0.56 0.67 0.33 1 0.56 0.78 FR (0.12) 0.33 0.44 0.78 0.89 0.33 0.44 0.56 0.67 0.56 1 0.56 0.78 SCRM in Indian Manufacturing Industries 137 RP RM RA RS DSR 0.04 0.1 0.04 0.07 0.06 0.12 0.06 0.09 0.07 0.13 0.07 0.1 SSR 0.05 0.11 0.02 0.08 0.06 0.12 0.03 0.09 0.08 0.14 0.05 0.11 LR 0.01 0.1 0.01 0.07 0.03 0.12 0.03 0.09 0.04 0.13 0.04 0.1 Table 25 Weighted Fuzzy Decision matrix for alternatives RLB 0.01 0.1 0.01 0.04 0.03 0.12 0.03 0.06 0.04 0.13 0.04 0.07 IR 0.04 0.09 0.07 0.07 0.05 0.11 0.08 0.08 0.07 0.12 0.09 0.09 SDM 0.01 0.09 0.01 0.07 0.03 0.11 0.03 0.08 0.04 0.12 0.04 0.09 ER 0.01 0.09 0.04 0.06 0.02 0.1 0.05 0.07 0.04 0.11 0.06 0.09 FR 0.04 0.09 0.04 0.07 0.05 0.11 0.05 0.08 0.07 0.12 0.07 0.09 138 V. Viswanath Shenoi et al. DSR SSR LR RLB IR SDM ER FR FPIS 0.1 0.12 0.13 0.11 0.12 0.14 0.1 0.12 0.13 0.1 0.12 0.13 0.09 0.11 0.12 0.09 0.11 0.12 0.09 0.1 0.11 0.09 0.11 0.12 FNIS 0.04 0.06 0.07 0.02 0.03 0.05 0.01 0.03 0.04 0.01 0.03 0.04 0.04 0.05 0.07 0.01 0.03 0.04 0.01 0.02 0.04 0.04 0.05 0.07 Table 26 Fuzzy Positive and Negative ideal solution SCRM in Indian Manufacturing Industries 139 140 Table 27 Distance of alternatives from FPIS and FNIS, closeness coefficient Table 28 Rank of the alternatives with respect to their closeness coefficient in FTOPSIS V. Viswanath Shenoi et al. Alternatives RP RM RA RS Alternatives RM RS RA RP D+ 0.55 0 0.53 0.25 CC 1 0.57 0.09 0.05 D− 0.03 0.58 0.05 0.33 CC 0.05 1 0.09 0.57 Ranking 1 2 3 4 mediating effect. It is observed that if the p-value falls below the established α level of 0.05, indicating the association between the pair of IV and DV (i.e., individual risk construct and performance), risk is significantly reduced by the inclusion of the mediating variable (namely, risk planning, risk monitoring, risk avoidance, risk sharing). From the results obtained, it is inferred that the mitigation strategies have significantly reduced the risk. Hence, all hypotheses H1 through H4 are accepted. Inference from the Empirical Study We identify the constructs of SCRM through empirical study. The validation of constructs is done through a series of tests to identify the various aspects of SCRM. The present work reinforces that the organization needs to have a longterm association with the supply chain partners. The long-term associations between the trading partners need to be established by sharing of information, risk and rewards equitably. Our findings indicate that the mediators (namely, risk planning, risk monitoring, risk avoidance, and risk sharing) have mediating effect on the relationship between the constructs of SCRM, mitigation strategies, and the DV, taken one at a time. Implications for Practitioners In a competitive environment, the risk is inevitable. A strategic plan is necessary to face risk. In most cases, we observe that mediating variable, risk monitoring, significantly reduces the risk, followed by risk sharing, risk avoidance, and risk planning respectively. The findings suggest that regular monitoring of the supply chain is essential for a resilient supply chain. SCRM in Indian Manufacturing Industries 141 Suggestions and Recommendations for Policy Makers The policymakers should train the practicing managers in risk management. The process of SCRM involves the design of strategy and framework. An instrument is essential to validate the framework pertaining to the industry. Based on the framework, functional managers implement operational decisions. Managers in all the levels of the supply chain need to know their business and be aware of types and levels of risks perceived. Risk awareness leads to efficient monitoring of the supply chain and enables the successful implementation of the SCRM. 5.2 Results of Application of FCM to SCRM Instantaneous vector Q1 represents the risk perceived during an instant of time. The instantaneous vector Q1 is passed on to the adjacency represented in Table 3. The resultant vector is threshold to one and is passed on to the adjacency matrix until the equilibrium. Equilibrium is attained in the third iteration (refer to Table 29). The equilibrium vector obtained is passed on to the relationship matrix M . The results in Table 30 suggest that the risks are mitigated, and the impact of risks is alleviated to a greater extent, as the results show negative values or zeroes. The equilibrium vector obtained is passed on to the relationship matrix M. The results in Table 31 suggest Table 29 Results of product of instantaneous vector Q1 and dynamical system matrix E Q1 RS1 Q2 =˜ RS1 RS2 Q3 =˜ RS2 RS3 QDSR 1 0 1 8 1 8 QSSR 0 1 1 7 1 8 QLR 0 1 1 7 1 8 QRLB 0 1 1 7 1 8 QIR 0 1 1 7 1 8 QSDM 0 1 1 7 1 8 QER 0 1 1 7 1 8 QFR 0 0 0 8 1 8 R1 0 1 1 6 1 7 R2 0 1 1 4 1 5 Table 30 Results of product of equilibrium vector Q3 and relationship matrix M RS3 QDSR 1 QRP 0 QSSR 1 QRM 0 QLR 1 QRA 0 QRLB 1 QRS 0 QIR 1 QSDM 1 QER 1 QFR 1 R1 1 R2 1 Table 31 Results of product of instantaneous vector Q1 with relationship matrix M RS3 QDSR 1 QRP 5 QSSR 1 QRM 7 QLR 1 QRA 6 QRLB 1 QRS 6 QIR 1 QSDM 1 QER 1 QFR 1 142 V. Viswanath Shenoi et al. the levels at which the individual mitigation strategy needs to be implemented to mitigate the risk and enhance the performance. The highest score of 7 is obtained for risk monitoring, suggesting that regular monitoring of the supply chain is essential for making proactive decisions. The next highest score of 6 is realized for risk avoidance, and risk sharing and the score of 5 for risk planning. 5.3 Comparison of the Proposed FCM and Fuzzy TOPSIS Model The problem of selecting the best mitigation strategy is evaluated through two approaches. The proposed model is based on the decision makers’ perception of risk. The ranking of the mitigation strategy is comparatively the same using both the approaches (refer to Tables 28 and 31). Both the methods scale well in the selection of mitigation strategies in this case. The FCM approach interacts with every other risk factors before attaining the equilibrium. On the other hand, FTOPSIS helps us to evaluate the ranking of the alternatives to mitigate the risk factors. The FCM approach not only suggests the level of application of alternatives but also predicts the risk faced by the firm using the instantaneous vector (refer to Table 30). 6 Conclusions The study is performed to identify the constructs of SCRM from the manufacturers’ perspective. Validation of constructs is done through a series of tests to capture the various aspects of SCRM. The present work reinforces that the organizations needs to have a long-term association with the partners of the supply chain. It is evident that every firm encounters risks that arise from various entities of the supply chain. Hence, the supply chain managers need to be proactive in identifying the risk as and when it surfaces in the supply chain. If ignored, it may lead to grave consequences. Practicing managers can use the proposed methodology to identify the risk, which also suggests an appropriate mitigation strategy that reduces the impact of risk on the supply chain, resulting in improved performance. FCM approach predicts the future risk and also suggests the mitigation strategies. FTOPSIS ranks the alternate mitigation strategies. From our study, it is concluded that risks have to be identified as soon as they surface. In our study, both the approaches scale well in finding the appropriate mitigation strategies. Accordingly, one needs to implement an appropriate mitigation strategy to reduce the impact of risk and improve the supply chain performance. Our findings emphasize that monitoring supply chain is essential for sustained performance. Also, training of managers is essential to be risk-sensitive. Broader insights into the models are SCRM in Indian Manufacturing Industries 143 possible only through a larger data set. Due to operational constraints in the present study, the survey is restricted to the samples available. Future studies can also look into specific risk aspects based on different types of manufacturing and service industries. Acknowledgments The authors are most grateful to the reviewers of the questionnaire, who have suggested constructive improvements and augmentative comments. The authors also thank the respondents for taking their valuable time in participating in the empirical study. The authors are grateful to the editors and reviewers, who have given us a chance to improve the original version of the book chapter. References Al Zubayer, M. A., Ali, S. M., & Kabir, G. (2019). Analysis of supply chain risk in the ceramic industry using the topsis method under a fuzzy environment. Journal of Modelling in Management, 14, 792–815. Arcelus, F. J., Kumar, S., & Srinivasan, G. (2012). Risk tolerance and a retailer’s pricing and ordering policies within a newsvendor framework. Omega, 40(2), 188–198. Chen, J., Sohal, A. S., & Prajogo, D. I. (2013). Supply chain operational risk mitigation: a collaborative approach. International Journal of Production Research, 51(7), 2186–2199. CRISIL. (2010). Rating methodology for manufacturing companies. CRISIL Rating Report (pp. 1– 4). Crocker, L., & Algina, J. (1986). Introduction to classical and modern test theory. Fort Worth: Harcourt Brace Jovanovich College Publishers, Orlando. Cronbach, L. J. (1951). Coefficient alpha and the internal structure of tests. Psychometrika, 16(3), 297–334. Dağdeviren, M., Yavuz, S., & Kılınç, N. (2009). Weapon selection using the ahp and topsis methods under fuzzy environment. Expert Systems with Applications, 36(4), 8143–8151. Dath, T. N. S., Rajendran, C., & Narashiman, K. (2009). A conceptual framework for supply chain managment with specifice reference to a developing economy. International Journal of Logistics Systems and Management, 5(5), 473–524. Dath, T. N. S., Rajendran, C., & Narashiman, K. (2010). An empirical study on supply chain management in india: The perspective of original equipment manufacturers and suppliers. European Journal of Industrial Engineering, 4(1), 2–39. Faisal, M., Banwet, D. K., & Shankar, R. (2007). Management of risk in supply chains: SCOR approach and Analytic Network Process. Supply Chain Forum: An International Journal, 8(2), 66–79. FCCI. (2013). FCCI report on power transmission. http://www.ficci.com/spdocument/20311/ power-transmission-report_270913.pdf. [Online; accessed 25 Dec 20014]. Ghadge, A., Dani, S., Chester, M., & Kalawsky, R. (2013). A systems approach for modelling supply chain risks. Supply Chain Management: An International Journal, 18(5), 523–538. Glykas, M. (2013). Fuzzy cognitive strategic maps in business process performance measurement. Expert Systems with Applications, 40(1), 1–14. Hair, J. F., Black, W. C., Babin, B. J., Anderson, R. E., & Tatham, R. L. (2006). Multivariate data analysis (Vol. 6). Upper Saddle River: Pearson Prentice Hall. Hendricks, K. B., & Singhal, V. R. (2005). Association between supply chain glitches and operating performance. Management Science, 51(5), 695–711. Kandasamy, W. B. V., Smarandache, F., & Ilanthenral, K. (2007). Elementary fuzzy matrix theory and fuzzy models for social scientists (Infinite Study). Los Angeles: Automaton. 144 V. Viswanath Shenoi et al. Kaplan, R., & Saccuzzo, D. (2012). Psychological testing: Principles, applications, and issues. Andover: Cengage Learning. Kaur, A., Kanda, A., & Deshmukh, S. G. (2006). A graph theoretic approach for supply chain coordination. International Journal of Logistics Systems and Management, 2(4), 321–341. Kaur, A., Kanda, A., & Deshmukh, S. G. (2007). Coordination in supply chains: An evaluation using fuzzy logic. Production Planning & Control, 18(5), 420–435. Khan, A., & Pillania, R. K. (2008). Strategic sourcing for supply chain agility and firms’ performance: A study of Indian manufacturing sector. Management Decision, 46(10), 1508– 1530. Kırılmaz, O., & Erol, S. (2017). A proactive approach to supply chain risk management: Shifting orders among suppliers to mitigate the supply side risks. Journal of Purchasing and Supply Management, 23(1), 54–65. Kosko, B. (1986). Fuzzy cognitive maps. International Journal of Man-Machine Studies, 24(1), 65–75. KPMG. (2010). KPMG Analysis Global Transport, Logistics in India, 2010. KPMG White Paper. Kristianto, Y., Gunasekaran, A., Helo, P., & Hao, Y. (2014). A model of resilient supply chain network design: A two-stage programming with fuzzy shortest path. Expert Systems with Applications, 41(1), 39–49. Kumar, N. (2005). Liberalisation, foreign direct investment flows and development: Indian experience in the 1990s. Economic and Political Weekly, 40, 1459–1469. Lavastre, O., Gunasekaran, A., & Spalanzani, A. (2012). Supply Chain Risk Management in French companies. Decision Support Systems, 52(4), 828–838. Liu, H., Chen, W., Kang, Z., Ngai, T., & Li, Y. (2005). Fuzzy multiple attribute decision making for evaluating aggregate risk in green manufacturing. Tsinghua Science & Technology, 10(5), 627–632. Manuj, I., Esper, T. L., & Stank, T. P. (2014). Supply chain risk management approaches under different conditions of risk. Journal of Business Logistics, 35(3), 241–258. Micheli, G. J. L., Mogre, R., & Perego, A. (2014). How to choose mitigation measures for supply chain risks. International Journal of Production Research, 52(1), 117–129. Moore, M., & Moore, M. K. (2003). A world without walls: Freedom, development, free trade and global governance. Cambridge: Cambridge University Press. NMCC. (2008). Current global financial crisis and its impact on India-measures for ensuring growth of the real sector of the economy. National Manufacturing Competitiveness Council Report (pp. 1–9). Oke, A., & Gopalakrishnan, M. (2009). Managing disruptions in supply chains: A case study of a retail supply chain. International Journal of Production Economics, 118(1), 168–174. Preacher, K. J. (2010). Calculation for the Sobel test: An interactive calculation tool for Mediation tests. http://quantpsy.org/sobel/sobel.htm. [Online; accessed 25 Dec 2014]. Preacher, K. J., & Hayes, A. F. (2008). Asymptotic and resampling strategies for assessing and comparing indirect effects in multiple mediator models. Behavior Research Methods, 40(3), 879–891. Radke, A. M., & Tseng, M. M. (2012). A risk management-based approach for inventory planning of engineering-to-order production. CIRP Annals-Manufacturing Technology, 61(1), 387–390. Rogers, H., Srivastava, M., Pawar, K., & Hartmann, E. (2012). A review of supply chain risk management research from an Indian perspective. In: Proceedings of European Risk Conference. Rostamzadeh, R., Ghorabaee, M. K., Govindan, K., Esmaeili, A., & Nobar, H. B. K. (2018). Evaluation of sustainable using an integrated fuzzy topsis-critic approach. Journal of Cleaner Production, 175, 651–669. Saeidi, P., Saeidi, S. P., Sofian, S., Saeidi, S. P., Nilashi, M., & Mardani, A. (2019). The impact of enterprise risk management on competitive advantage by moderating role of information technology. Computer Standards & Interfaces, 63, 67–82. Sawik, T. (2013). Selection of resilient supply portfolio under disruption risks. Omega, 41(2), 259–269. SCRM in Indian Manufacturing Industries 145 Shaw, K., Shankar, R., Yadav, S. S., & Thakur, L. S. (2012). Supplier selection using fuzzy AHP and fuzzy multi-objective linear programming for developing low carbon supply chain. Expert Systems with Applications, 39(9), 8182–8192. Shemshadi, A., Toreihi, M., & Shirazi, H. (2011). Supplier selection based on supplier risk: An anp and fuzzy topsis approach. Journal of Mathematics and Computer Science, 2, 111–121 Shenoi, V. V., Dath, T. N. S., & Rajendran, C. (2016). Supply chain risk management in the Indian manufacturing context: A conceptual framework. International Journal of Logistics and Systems Management, 25(3), 313–335. Sodhi, M. S., & Chopra, S. (2004). Avoiding supply chain breakdown. MIT Sloan Management Review, 46(1), 53–62. Stylios, C. D., Georgopoulos, V. C., Malandraki, G. A., & Chouliara, S. (2008). Fuzzy cognitive map architectures for medical decision support systems. Applied Soft Computing, 8(3), 1243– 1251. Sureshchandar, G. S., Rajendran, C., & Anantharaman, R. N. (2001). A holistic model for total quality service. International Journal of Service Industry Management, 12(4), 378–412. Tang, O., & Nurmaya Musa, S. (2011). Identifying risk issues and research advancements in supply chain risk management. International Journal of Production Economics, 133(1), 25–34. Tarei, P. K., Thakkar, J. J., & Nag, B. (2020). Benchmarking the relationship between supply chain risk mitigation strategies and practices: An integrated approach. Benchmarking: An International Journal, 27, 1683–1715. Thun, J. H., & Hoenig, D. (2011). An empirical analysis of supply chain risk management in the German automotive industry. International Journal of Production Economics, 131(1), 242– 249. Trkman, P., & McCormack, K. (2009). Supply chain risk in turbulent environments a conceptual model for managing supply chain network risk. International Journal of Production Economics, 119(2), 247–258. Wagner, S. M., & Bode, C. (2006). An empirically-based and more fine-grained investigation into supply chain risks. In Proceedings of 4th Worldwide Research Symposium on Purchasing and Supply Chain Management, San Diego. Wagner, S. M., & Bode, C. (2008). An empirical examination of supply chain performance along several dimensions of risk. Journal of Business Logistics, 29(1), 307–325. Zadeh, L. A. (1965). Fuzzy sets. Information and Control, 8(3), 338–353. Improving Service Supply Chain of Internet Services by Analyzing Online Customer Reviews Suchithra Rajendran and John Fennewald 1 Introduction Service supply chain, a series of activities involving the flow of service, includes customers, service providers and service producers, as the main entities (Sampson 2000). Among these components, customers and their opinion are considered to be essential factors in the production process (Shahin 2010). In particular, services, such as wireless telecommunications and internet service providers, involves direct customer-provider interaction, and hence, it is necessary to investigate such services for quality improvement. More specifically, the wireless communications industry is proliferating with a reported growth of 4.1% from 2013 to 2018 and an expected growth of 3.5% from 2018 to 2023 in the United States (Ibis World Report 2020). This expected industry growth will lead to an estimated revenue increase of $114.6 billion (Ibis World Report 2020). Several components factor into this anticipated advancement, including faster connectivity speeds and better service, which leads to an overall desire for increased internet usage and the idea that internet access is a necessity rather than viewing it was a luxury (Ibis World Report 2020). The wireless communications industry is highly competitive, as approximately two-thirds of the adults use broadband internet at home (Kohut 2018). Despite this level of competition, in the United States, the industry is dominated by very few noticeable players, as the five biggest companies combine to control over 68% of total market share—AT&T (24.9%), Comcast (17.2%), Charter S. Rajendran () · J. Fennewald Department of Industrial and Manufacturing Systems Engineering, University of Missouri, Columbia, MO, USA Department of Marketing, University of Missouri, Columbia, MO, USA e-mail: RajendranS@missouri.edu © Springer Nature Switzerland AG 2021 S. Srinivas et al. (eds.), Supply Chain Management in Manufacturing and Service Systems, International Series in Operations Research & Management Science 304, https://doi.org/10.1007/978-3-030-69265-0_5 147 148 S. Rajendran and J. Fennewald (12.9%), Verizon (7.4%), and CenturyLink (5.8%) (Ibis World Report 2020). In such a competitive environment and a quickly growing industry, Internet Service Providers (ISPs) must ensure that their efforts are targeted towards attracting and retaining customers for the continued growth of their business. Nearly 78% of revenue originates in the ISP-to-customer sector, representing the largest market in which ISPs operate and compete. This sector includes both residential and business services with the most common activities involving downloading multimedia contents, browsing the internet, and using basic internet applications such as accessing email accounts (Ibis World Report 2020). ISPs in this domain compete on many factors, namely, availability of service, price, speed of connectivity, and customer service (Ibis World Report 2020). In order to attract and retain the most customers possible, ISPs must determine the client’s expectations and tailor their offerings to meet consumer needs and wants. One of the best ways for ISPs to identify what their users perceive about their services is to analyze online reviews posted by customers. Several social networking sites allow consumers to discuss their experiences with various ISPs and detail their levels of satisfaction or dissatisfaction. As Web 2.0 has gained traction and popularity, this style of content that is generated by users has increased rapidly (Guo et al. 2017). Reviews can be either negative or positive, and since negative reviews circulate five to six times faster than positive reviews, their impact on the business can be monumental (Salehan and Kim 2016). As social media usage continues to grow, enterprises experience additional roadblocks since they must examine OCR to understand how customers are interacting with their products and to determine the performance of their products and services (Sen and Lerman 2007). Recently, studies have also investigated online customer reviews (OCR) to find the strengths and weaknesses associated with the products and services offered by a business (Salehan and Kim 2016; Srinivas and Rajendran 2019; Rajendran 2020). By using OCR to understand the strengths and weaknesses of their offerings, any organization, including ISPs, can discover more ways to improve and enhance the experiences that customers have with their products and services. Insights that ISPs may extract from online customer reviews include deciding which of their services to improve in order to satisfy their clients who have been disappointed in the past or determining the areas in which they perform very well and developing marketing plans to highlight those fields of service to attract more customers (Rajendran and Pagel 2020a). As studies, such as Cheung and Lee (2012), have shown that new customers place heavy emphasis on the opinions of existing clients and that OCR weigh heavily into a consumer’s view on which company to choose, it is extremely likely that potential ISP consumers might check OCR to gain insight into the experiences that prior users have had with each ISP that they are considering. Each ISP adds value to customers in a different way tailored to its own business model. Many companies find success through economies of scale by bundling phone, internet, and television services into one package for consumers. Other ISPs find a niche market and focus on providing the best service in a particular sector, Improving Service Supply Chain of Internet Services by Analyzing Online. . . 149 such as rural areas (Ibis World Report 2020). Regardless of what an ISP’s business model is, OCR carry abundant information that can be beneficial to ISPs as they seek to provide the most value for their customers. The purpose of this paper is to identify the strengths, weaknesses, opportunities, and threats of an ISP utilizing knowledge mining to improve customer satisfaction based on information gained from online consumer reviews. Each consumer review explores various topics (e.g., speed, connection, customer service, etc.), and conveys an attitude (positive, negative, or neutral) towards each topic. In this study, we use text analytic methods to extract and analyze thousands of reviews for identifying those key topics and propose managerial recommendations based on our results. The rest of the paper is organized as follows. Section “Literature Review” covers a review of several prominent studies discussed in the literature. A thorough discussion of the proposed methodological approach is given in Section “Methodology”. Section “Case Study” includes the description of the case study as well as an analysis of the results obtained. Lastly, Section “Conclusions” presents the conclusions and the scope of later work. 2 Literature Review As discussed earlier, the focus of this paper is to capture the voice of customers using online review mining and enable ISPs to provide better service. Therefore, in this section, we present some prominent studies related to wireless service providers and the impact of word of mouth on companies. 2.1 Wireless Service Providers One of the very first studies on customer expectations in the wireless service industry was pioneered by Zander (1997). The author explored the idea that customers of wireless communication services expect to receive the same services as they would receive from a fixed network. The paper analyzed both a “universal” coverage scenario providing wideband services at all locations and a “hot-spot” scenario where coverage is offered, but only in select geographical areas. The study showed companies the benefits of presenting wireless services to all consumers and but also indicated that as bandwidth increases, the infrastructure cost of a wireless system would also increase. Bar and Park (2005) looked at public Wi-Fi networks in the United States and the policy issues that the new trends raise. Metropolitan cities already have the existing infrastructure which supports the Wi-Fi networks, and they can help provide connectivity for city employees while making urban areas more desirable for residents and businesses. With the rapid development in technology, broadband connectivity has become more widespread, and voice services are more mobile; as a result, ISPs are bundling 150 S. Rajendran and J. Fennewald voice services with broadband connectivity in homes. According to Markendahl and Makitalo (2007), ISPs consider the following important factors: network operation and access provisioning, customer acquisition and relations, trustworthy relationship with the customer. The authors also hypothesize that actors from outside the telecommunication industry will venture into this business and that wireless access is increasingly becoming a service with a wide coverage area rather than existing solely in the location where the service is officially provided. Seo et al. (2008) highlighted the challenges that the wireless telecommunication market is facing, especially with the growing competition. Unlike goods, wireless services have to be offered continuously, and hence, developing strategies to retain and attract new customers is essential. Their study tried to understand how variables such as customer-related factors (e.g., age and gender) affect service plan complexity. A similar research was conducted by Quach et al. (2016), in which the authors examined the different aspects of an ISP’s service quality and their impact on customer loyalty. They collected data from more than 1200 internet users, segmented the information based on the customer’s internet usage, and recorded their perception of service quality dimensions. Their study proved that the perceived service quality dimension has an impact on both attitudinal and behavioral loyalty. 2.2 Electronic Word of Mouth (eWOM) Erkan and Evans (2016) examined the influence of eWOM in social media on consumers’ purchasing intentions. Several hypotheses were considered, and the authors proved that eWOM is positively related to consumers’ purchasing intention, the usefulness of eWOM information is positively associated with the adoption of eWOM information, and the quality of eWOM is positively associated with usefulness. Wang et al. (2016) showed several differences between traditional WOM and eWOM, including the anonymity of eWOM messages, the fact that multiple people can access eWOM at any time, and that eWOM is more persistent and measurable. Their research indicated that about 70% of 28,000 internet users in 56 countries rely on online consumer reviews for recommendations. Text analytic tools for examining eWOM have not only been used in product improvement, but also in the service sector (Rajendran and Pagel 2020a). For instance, Liang et al. (2018) studied the effect of eWOM on hospitality and tourism management. Their study showed that eWOM is highly influential due to its speed, amplitude, and convenience. This idea, coupled with the fact that WOM reduces a company’s ability to influence consumers through traditional marketing and advertising, shows how vital it is for companies to become actively involved in online consumer communities (Rajendran and Pagel 2020b). While the studies mentioned above examined the relationship between online reviews and customer purchasing intentions, Jalilvand et al. (2011) investigated the reason for the extensive spread of WOM. Their research showed that consumers spread WOM for several reasons, including extreme satisfaction or dissatisfaction, Improving Service Supply Chain of Internet Services by Analyzing Online. . . 151 commitment to the firm, length of the relationship with the firm, and novelty of the product. Additionally, users may be motivated to share their experiences due to satisfaction, pleasure, or sadness. The authors showed that the main antecedents of WOM influence are demographic similarity, and perceptual affinity, and that WOM could influence product evaluations. Balaji et al. (2016) examined how negative WOM (NWOM) on social networking sites (SNS) can adversely impact companies. Customer complaints can lead to brand dilution, volatile stock prices, and on a grand scale can even be the cause for public relations crises. Approximately 77% of online shoppers rely on user reviews to make purchasing decisions, and over one million people read product or service reviews every week on SNS, and more than 80% of these reviews are negative. Their study looked at the cognitive dissonance and social support theories to understand the determinants of NWOM on SNS. Since sharing opinions is a social activity, the social support theory shows that NWOM sharing may be due to the desire for social interaction. Additionally, consumers may engage in NWOM communications in order to reduce their cognitive dissonance levels because they are frustrated with a product or service. This proposed research is one of the first to analyze online customer reviews posted for internet service providers using text analytics. We propose insights and managerial recommendations based on strategic planning tool (SWOT) and six sigma techniques (root cause analysis) to capture the voice of the customers. Though the literature discusses few studies that are conducting the SWOT analysis for internet service providers, to the best of the authors’ knowledge, this study is the first to leverage customers’ reviews and understand the voice of the customers for this purpose. 3 Methodology Figure 1 presents a flow diagram showing the methodological framework used for data collection and analysis. The proposed approach will first use a web scraper to extract thousands of publicly available customer review archives from online sources. Next, the reviews are separated into three categories based on the consumer’s rating (which is on a scale of 1–5). On the rating scale, a value of “1” indicates a highly-dissatisfied customer, and “5” indicates a highly-satisfied customer. All reviews with a rating of “1” or “2” are marked as “negative”, reviews with a score of “3” are labeled as “neutral”, and reviews with a rating of “4” or “5” are labeled as “positive”. Upon separation into negative, neutral, and positive categories, bigrams and trigram analyses are used to identify the bag of words commonly cooccurring and gain further insights into the context in which consumers use certain terms. These frequently co-occurring words are analyzed and compared across different companies in the ISP industry with the goal of creating a customer-centric SWOT and a root cause analysis using the negative, neutral, and positive bigrams and trigrams. Figure 1 provides an overview of the proposed approach. 152 S. Rajendran and J. Fennewald Fig. 1 Methodological Framework 3.1 Web Scraping and Text Pre-Processing The consumer reviews of the ISPs under study are scraped from several online customer review sites, and since the customer feedback might not be particularly centered around any one single aspect of the corporation, reviews are separated into individual sentences, which are analyzed as independent comments. The reviews are cleaned prior to doing the text analysis by removing incomplete, duplicate, and irrelevant reviews. Subsequently, the sentences are tokenized, special characters and non-English words are eliminated, inflected words are stemmed, characters are converted to lowercase, and stop and infrequent words are removed. 3.2 Topic Identification Using Bigrams and Trigrams Analyses After scraping and pre-processing, the most frequent topics discussed in reviews need to be identified, which is achieved by analyzing the frequent words, bigrams, and trigrams. A bigram is a combination of two words that are used consecutively in Improving Service Supply Chain of Internet Services by Analyzing Online. . . 153 relevance to the same topic. Examples pertaining to our study would be “customer service” or “connectivity issues”. A trigram is a group of three words that are used consecutively, and examples would be “good customer service” and “connectivity issues rural”. This section presents an analysis of bigrams and trigrams, as discussed by Jurafsky and Martin (2014). If x1 , x2 , . . . , xN are a set of words, then the probability that these words occur in a sequence is denoted by P(X) as shown in Eq. (1). P(X) = P (x1 . . . xN ) (1) The probability of x2 occurring after x1 in a sequence is shown in Eqs. (2) and (3). Similarly, the probability of x3 occurring after x2 and x1 in a sequence is given in Eqs. (4) and (5) is derived from Eq. (4). P (x2 |x1 ) = P (x1 , x2 ) P (x1 ) (2) P (x1 , x2 ) = P (x2 |x1) × P (x1 ) (3) P (x1 , x2 , x3 ) = P (x3 |x1 , x2 ) × P (x1 , x2 ) (4) P (x1 , x2 , x3 ) = P (x1 ) × P (x2 |x1 ) × P (x3 |x1 , x2 ) (5) If we extrapolate this expression to N words, the probability of word xN occurring in a word sequence x1 , x2 , . . . , xN − 1 is given by Eqs. (6) and (7). P (x1 . . . xN ) = P (x1 ) × P (x2 |x1 ) × P (x3 |x1 , x2 ) . . . P (xN |x1 , x2 . . . xN−1 ) (6) P (x1 . . . xN ) = N Pr (xn |x1 . . . xn−1 ) (7) n=1 Applying the conditional probability chain rule to the sequences of words under consideration, Eqs. (8) and (9) give the probability of N words occurring in a sequence. P (x1 . . . xN ) = P (x1 ) × P (x2 |x1 ) × P x3 |x12 . . . P xN |x1N−1 P (x1 . . . xN ) = N P xN |x1N−1 (8) (9) i=1 In Eq. (10), we consider the Markovian assumption for the bigram model. P xN |x1N−1 = P (xN |xN−1 ) (10) 154 S. Rajendran and J. Fennewald where P(xN | xN − 1 ) is approximated by using the proportion of the bigram count of xN − 1 and xN (represented by ν(xN − 1 xN )) to the sum of the frequency of all the bigrams containing xN − 1 ( x ν (xN−1 x)) as shown by Eq. (11). P (xN |xN−1 ) = ν (xN−1 xN ) x ν (xN−1 x) (11) Constraint (12) uses the trigram model to predict the conditional probability of the Nth word in a sequence. P (xN |xN−1 , xN−2 ) = ν (xN−2 xN−1 xN ) x ν (xN−2 xN−1 x) (12) 3.3 SWOT Strategic Planning Using Bigrams and Trigrams Analyses For each ISP, the reviews are separated based on their ratings as negative (≤2 stars), neutral (3-stars), and positive (≥4 stars), following which commonly co-occurring bigrams and trigrams are identified for each category. These bigrams and trigrams are then used to conduct a SWOT analysis for individual ISPs and the industry as a whole. SWOT is a planning tool most commonly used for recognizing the strategic factors (both internal and external) that are important to the development of an enterprise (Kurttila et al. 2000). The strengths and weaknesses are identified by analyzing the internal characteristics of the organization, while opportunities and threats constitute the external elements, such as competition. The basic framework of SWOT analysis is shown in Fig. 2. Such a SWOT assessment highlights the Internal Weaknesses ISP offerings that enrich customer’s Qualities of the ISP that hurt customer experience and lead to positive WOM experience and create a negative (e.g., fast speeds, helpful technicians) perception (e.g., poor customer service, slow speeds in rural areas) External Strengths Opportunities Potential changes companies can adapt to enhance reputation (e.g., bundling packages, better trained customer service representatives) Threats Obstacles that the ISP faces due to external environment (e.g., governmental regulation, competition from other companies) Fig. 2 Basic Framework of SWOT Analysis based on ISP Customer Experience Improving Service Supply Chain of Internet Services by Analyzing Online. . . Table 1 Overview of the ISPs considered Company ISP#1 ISP#2 ISP#3 ISP#4 ISP#5 ISP#6 ISP#7 ISP#8 Primary services offered Wireless Internet Phone services x x x x x x x x x x x 155 Television x x critical factors that influence the viewpoint of the corporation and can aid the ISP in their planning decisions. 4 Case Study In this study, we have chosen eight companies that control over 70% of the total market share in the ISP domain. We have de-identified and referred to the ISPs under consideration as ISP#1 through #8. In total, 23,145 online customer reviews are considered with the following breakdown of reviews for each company: ISP#12407, ISP#2-4208, ISP#3-937, ISP#4-3446, ISP#5-989, ISP#6-3324, ISP#7-5019 and ISP#8-2815. Table 1 shows an overview of the ISPs considered and the primary services offered. As it can be seen, wireless internet is a primary service for every company except ISP#3 and ISP#6; and ISP#2 and ISP#3 are the only two companies that offer entertainment media as their primary services. 4.1 Data Description Table 2 shows the data features and an explanation of each field. The two fields that are considered in this study are “Rating” and “Review”. Figure 3 shows the distribution of reviews for the different internet service providers under consideration. We can see that the majority of the time, all companies are being negatively perceived by customers. More specifically, over 80% of the total reviews that are posted for ISP#1, ISP#3, ISP#4 and ISP#8 are given less than a 2-star rating. On the other hand, relatively more positive reviews are observed for ISP#5, whereas neutral reviews are comparatively more in ISP#6. 156 S. Rajendran and J. Fennewald Table 2 Data fields Field Rating Review Explanation Contains a rating on a scale from 1–5 of how much the customer liked the service. The value “1” represents a highly unsatisfied customer while a “5” represents a highly satisfied customer Contains the review body Fig. 3 Distribution of the Reviews 4.2 Experimental Results Once the reviews are extracted, the data is cleaned using the procedure discussed in Sect. “Web Scraping and Text Pre-Processing”. This data pre-processing is done with the modules in the Python natural language toolkit (NLTK). In addition to the default English stop words in the NLTK toolkit, we also created a custom set of stop words to filter unnecessary words that do not add value to our study. Table 3 gives an overview of the topics that occur most frequently in the OCR, a description for each topic, and the significant words associated with each topic. Clearly, ‘customer service’, ‘connectivity’ and ‘speed’ are the key topics that customer reviews are emphasized on. For the purpose of illustration, the most frequently occurring bigrams and trigrams related to the ‘service’ Topic is shown in Fig. 4. Many of these word combinations revolve around customer service or whether the actual services (wireless internet, TV, phone) are good or bad. Table 4 shows the list of most positively- and negatively-discussed topics based on bigram and trigram analyses. Topics related to customer service, connectivity, compatibility with other devices, billing and payments are several of the most common. Improving Service Supply Chain of Internet Services by Analyzing Online. . . 157 Table 3 Key topics overview Topic Service Phone Contract Rural area Representative Speed Description The service provided to customers by their ISP Use of phone with service from an ISP The signed agreement between ISPs and their customers ISP performance in sparsely populated areas Employee of an ISP who helps customers with difficulties Measure of how quickly service reaches customers Significant words Slow installation, setup time Coverage, connectivity, cell, data Cancel, terminate, switch, billing Rural connection, dead zone House calls, technician, unhelpful Fast, slow, download, high BIGRAMS (‘customer’, ‘service’), (‘cancel’, ‘service’), (‘phone’, ‘service’), (‘service’, ‘call’), (‘service’, ‘slow’), (‘tv’, ‘service’) TRIGRAMS (‘worst’, ‘customer’, ‘service’), (‘terrible’, ‘customer’, ‘service’), (‘contact’, ‘customer’, ‘service’), (‘poor’, ‘customer’, ‘service’), (‘horrible’, customer’, ‘service’), (‘call’, ‘customer’, ‘service’), (‘customer’, ‘service’, ‘rep’), (‘call’, ‘cancel’, ‘service’) Fig. 4 Frequently Occurring Bi-grams and Tri-grams for Negative Opinions Related to the Topic ‘Service’ 4.3 SWOT Analysis From the topics discussed in Table 4, SWOT analysis for each ISP can be developed. A sample SWOT Analysis for ISP#1 is given in Table 5. Following that, we also propose several managerial insights based on the observations made. The managerial insights derived from the SWOT analysis pertaining to a specific company (ISP#1) is presented below. A similar approach can be adopted for other providers as well. • Customized Bundle Plans: Develop a bundling plan that can be tailored to individual consumer’s preferences. In other words, the most effective package will be the one which allows consumers to pay for specific services which they value and nothing more (e.g., fast speeds to play games online but limited TV channels, wide range of TV channels without high-speed internet). • Employee Training: Ensure that all customer service representatives are well trained on how to best serve their customer base. This may include (i) educating them on products, policies, and packages that the company offers, and (ii) clearly communicating with customers on billing dates and plans. 158 S. Rajendran and J. Fennewald Table 4 Positively and negatively perceived topics by customers for each ISP ISP ISP#1 Topics based on positive reviews Fast speeds internet in cities; versatility; easy installation ISP#2 Helpful technician; good streaming for playing online games Triple play package; fast speed; wide TV channel variety Works well even during bad weather; good upload and download speeds ISP#3 ISP#4 ISP#5 ISP#6 ISP#7 ISP#8 Good family plan; cost effective; service representative speaks in English fluently Binge data streaming; good service in Europe; service representatives give good instructions Consistent internet service; helpful salespeople Helpful technicians; wide TV channel selection; easy and quick troubleshoots; providing house calls services Topics based on negative reviews Slow service in rural areas; changing billing dates and switching plans without prior notification; routers are not durable; slow speed while streaming media Wrongful billing; increasing service costs Hidden fees; disrupted TV streaming Slow installation; overseas call centers; fixed 2-year contract; poor transition from one generation to its next Unauthorized payments and charges; hard to cancel contract Offshore call centers; incompatible with certain android devices Poor international service; call center representatives do not speak English fluently; poor signal in rural areas; slow 4G LTE Services are priced high; DVR is expensive Table 5 SWOT Analysis for ISP#1 Strengths – Fast speeds – Quick installation – Widespread coverage – Highly compatible across different devices Opportunities – Expand TV channel offering – More comprehensive training for customer service representatives – Improve services in rural areas Weaknesses – Routers are not durable – Irregular payments – Expensive – Switching plans and billing dates – Slow streaming speed of media-services Threats – Well marketed triple and quadruple play packages by other ISPs – House calls services offered by certain companies • Marketing Company Strengths: Market the company’s strongest suits, such as high-speed internet and widespread coverage. Moreover, as one of the industry leaders, it is also essential to advertise that ISP#1 possesses a large portion of the market share. Improving Service Supply Chain of Internet Services by Analyzing Online. . . 159 4.4 Link and Root Cause Analyses Figure 5 presents a positive link analysis representing the overall topics that are commonly discussed in the positive reviews across all ISPs. A larger-sized node indicates that the topic is spoken more often by customers. Clearly, we can see that ‘Bundling’, ‘Connectivity’, and ‘Technical Service’ were three of the most repeated topics in the positive review domain. Under the ‘bundling’ topic, customers praise the triple and quadruple play packages as well as the family plans due to their cost savings. The topic ‘Connectivity’ is observed to be positively spoken by customers of certain companies and negatively for other ISPs. Thus we interpret that connectivity is a key issue that companies must pay attention to. Customers prioritize consistent connection over speed, while good online gaming streaming is also a factor that they consider. During a technical issue, customers expect proper assistance either by visiting them at home or solving the problem over the phone (i.e., easy troubleshooting). Figure 6 depicts the results of the root cause analysis through a fishbone diagram highlighting the overall causes that lead to customer dissatisfaction in the internet service sector. The primary causes are ‘Contracts’, ‘Connectivity’, ‘Cell Phone Service’, and ‘Call Center’. The associated secondary causes are also shown in Fig. 6. Clearly, customers are highly unsatisfied when they are charged with additional fees without prior notification. Interrupted service and slow loading speed in rural areas are also some issues that users experience. Moreover, they are not comfortable with automation, especially studies have proven that the older population is resistant to automation changes (Vaportzis et al. 2017). Fig. 5 Positive link analysis 160 S. Rajendran and J. Fennewald Fig. 6 Root cause analysis Ishikawa Diagram 4.5 Managerial Implications Based on the bigram and trigram analyses, we suggest the following recommendations to the internet service providers as a whole. • Streamline customer service call process: When customers attempt to contact ISPs over the phone, they often complain that it is difficult to speak to a human and that processes are automated. Our recommendation would be for ISPs to have online live chat programs that can achieve a balance between customer satisfaction and cost. • Provide positive work environment for customer service representatives: Many online reviews mention frustrations with customer service representatives, specifically for being “rude”, “unhelpful” or “not caring” about the user problems. Prior studies have proven a positive relationship between employee job stress, job satisfaction, and customer loyalty (Loveman 1998; Hansemark and Albinsson 2004; Hill and Alexander 2017). Therefore, we theorize that the rude behavior of customer service agents is due to the burnout that these employees experienced from answering customers’ calls monotonously. Hence it is recommended that ISPs provide a more positive work environment offering free/discounted health club memberships and conduct regular medical checkups. Also, ISPs can develop a handbook for call-center representatives providing a set on instructions on how to handle different types of customer grievances. • Payment communication: Many consumers are upset that ISPs did not communicate clearly about contracts, extra fees, change in billing dates, etc. Our recommendation for ISPs is to interact and update customers by different modes of communication (such as, phone calls, emails, and text messages). Improving Service Supply Chain of Internet Services by Analyzing Online. . . 161 • Rural area connectivity: A large portion of customers complain about poor internet service in rural areas. We suggest that stronger infrastructure has to be established in the countryside to better serve their clients and provide faster speeds. • Device compatibility: Some consumers criticize that their internet service does not work with various devices, whether that be their brand of routers, cellphones, or their gaming console. To avoid this issue, we recommend that ISPs keep device compatibility as a priority when designing their products and services. • Improve customer service representative language training: Based on the reviews, consumers are commonly distressed that the representatives lack English fluency and hence are not able to entirely comprehend their grievances. We recommend that ISPs offer extensive language training for their representatives who are non-native English speakers. 5 Conclusions In recent years, due to the increasing popularity of service sectors, service supply chains are being widely investigated to improve service quality. Studies have shown that online consumer reviews heavily influence consumer’s decisions across many industries, including ISPs. This research is one of the first to provide managerial insights for ISPs to improve the wireless services offered to their customers by analyzing online user reviews. While very few research has been conducted involving SWOT analysis from an internet provider’s perspective, examining online customer reviews enables ISPs to understand their consumer needs better. Over 23,000 online reviews were first collected using web scraping, reviews were subsequently separated based on their star ratings, following which commonly cooccurring words were identified in each category, and a SWOT and root cause analyses were finally conducted. We identified six meaningful topics using our approach and conducted a SWOT analysis to showcase the critical factors that impact consumers’ choice of ISP. We also proposed the following managerial recommendations to the ISP sector, based on the results obtained. • • • • • • Streamline customer service call process Provide a positive work environment for customer service representatives Interact with customers using different means of communication Establish proper infrastructure in rural areas to improve connectivity Make design changes to ensure device compatibility Provide language training for representatives 162 S. Rajendran and J. Fennewald References Balaji, M. S., Khong, K. W., & Chong, A. Y. L. (2016). Determinants of negative word-of-mouth communication using social networking sites. Information & Management, 53(4), 528–540. Bar, F., & Park, N. (2005). Municipal Wi-Fi networks: The goals, practices, and policy implications of the US case. Communications and Strategies, 61, 107–126. Cheung, C. M., & Lee, M. K. (2012). What drives consumers to spread electronic word of mouth in online consumer-opinion platforms. Decision Support Systems, 53(1), 218–225. Erkan, I., & Evans, C. (2016). The influence of eWOM in social media on consumers’ purchase intentions: An extended approach to information adoption. Computers in Human Behavior, 61, 47–55. Guo, Y., Barnes, S. J., & Jia, Q. (2017). Mining meaning from online ratings and reviews: Tourist satisfaction analysis using latent dirichlet allocation. Tourism Management, 59, 467–483. Hansemark, O. C., & Albinsson, M. (2004). Customer satisfaction and retention: The experiences of individual employees. Managing Service Quality: An International Journal, 14(1), 40–57. Hill, N., & Alexander, J. (2017). The handbook of customer satisfaction and loyalty measurement. New York: Routledge. Ibis World Report (2020). https://www.ibisworld.com/industry-statistics/market-size/wirelesstelecommunications-carriers-united-states/ Jalilvand, M. R., Esfahani, S. S., & Samiei, N. (2011). Electronic word-of-mouth: Challenges and opportunities. Procedia Computer Science, 3, 42–46. Jurafsky, D., & Martin, J. H. (2014). Speech and language processing (p. 3). London: Pearson. Kohut, A. (2018). Demographics of internet and home broadband usage in the United States. Pew Research Center: Internet, Science & Technology. Retrieved February 5, 2018, from www.pewinternet.org/fact-sheet/internet-broadband/ Kurttila, M., Pesonen, M., Kangas, J., & Kajanus, M. (2000). Utilizing the analytic hierarchy process (AHP) in SWOT analysis—A hybrid method and its application to a forest-certification case. Forest Policy and Economics, 1(1), 41–52. Liang, L. J., Choi, H. C., & Joppe, M. (2018). Exploring the relationship between satisfaction, trust and switching intention, repurchase intention in the context of Airbnb. International Journal of Hospitality Management, 69, 41–48. Loveman, G. W. (1998). Employee satisfaction, customer loyalty, and financial performance: An empirical examination of the service profit chain in retail banking. Journal of Service Research, 1(1), 18–31. Markendahl, J., & Makitalo, O. (2007, June). Analysis of business models and market players for local wireless internet access. In 6th Conference on Telecommunication Techno-Economics (pp. 1–8). IEEE. Quach, T. N., Thaichon, P., & Jebarajakirthy, C. (2016). Internet service providers’ service quality and its effect on customer loyalty of different usage patterns. Journal of Retailing and Consumer Services, 29, 104–113. Rajendran, S. (2020). Improving the performance of global courier & delivery services industry by analyzing the voice of customers and employees using text analytics. International Journal of Logistics Research and Applications, 2020, 1–21. Rajendran, S., & Pagel, E. (2020a). Recommendations for emerging air taxi network operations based on online review analysis of helicopter services. arXiv preprint arXiv:2006, 10898. Rajendran, S., & Pagel, E. (2020b). Insights on next-generation manufacturing of smart devices using text analytics. Heliyon, 6(7), e04491. Salehan, M., & Kim, D. J. (2016). Predicting the performance of online consumer reviews: A sentiment mining approach to big data analytics. Decision Support Systems, 81, 30–40. Sampson, S. E. (2000). Customer-supplier duality and bidirectional supply chains in service organizations. International Journal of Service Industry Management, 11, 17. Sen, S., & Lerman, D. (2007). Why are you telling me this? An examination into negative consumer reviews on the web. Journal of interactive marketing, 21(4), 76–94. Improving Service Supply Chain of Internet Services by Analyzing Online. . . 163 Seo, D., Ranganathan, C., & Babad, Y. (2008). Two-level model of customer retention in the US mobile telecommunications service market. Telecommunications Policy, 32(3–4), 182–196. Shahin, A. (2010). SSCM: Service supply chain management. International Journal of Logistics Systems and Management, 6(1), 60–75. Srinivas, S., & Rajendran, S. (2019). Topic-based knowledge mining of online student reviews for strategic planning in universities. Computers & Industrial Engineering, 128, 974–984. Vaportzis, E., Giatsi Clausen, M., & Gow, A. J. (2017). Older adults perceptions of technology and barriers to interacting with tablet computers: A focus group study. Frontiers in Psychology, 8, 1687. Wang, Y., Wallace, S. W., Shen, B., & Choi, T. M. (2015). Service supply chain management: A review of operational models. European Journal of Operational Research, 247(3), 685–698. Wang, T., Yeh, R. K. J., Chen, C., & Tsydypov, Z. (2016). What drives electronic word-of-mouth on social networking sites? Perspectives of social capital and self-determination. Telematics and Informatics, 33(4), 1034–1047. Zander, J. (1997, May). On the cost structure of future wideband wireless access. In IEEE 47th Vehicular Technology Conference. Technology in Motion (Vol. 3, pp. 1773–1776). IEEE. An Integrated Problem of Production Scheduling and Transportation in a Two-Stage Supply Chain with Carbon Emission Consideration Bobin Cherian Jos, Chandrasekharan Rajendran, and Hans Ziegler 1 Introduction Sequencing and scheduling involve decision-making that plays a vital role in various manufacturing and service industries (Pinedo 2012, pp. 1–5). In practice, many of these industries strive to ensure the effective use of available resources to meet the organization’s specific objectives by improving the efficiency in scheduling and sequencing. Supply chain management involves the reduction in operational costs, improving customer service levels, and establishing a competitive advantage by integrating and coordinating different functions of the supply chain (Chopra et al. 2016, pp. 1–10). The scheduling problem in a supply chain may face several constraints arising from various levels of the chain. Therefore, scheduling theory assumes great significance in integrated supply chain planning as well as logistic planning activities in organizations aiming to improve customer service (Azadian et al. 2015). In this study, an integrated logistics and scheduling problem with the consideration of carbon emissions in a supply chain system is described, and a mixed-integer linear programming (MILP) model is proposed. B. C. Jos () Department of Mechanical Engineering, Mar Athanasius College of Engineering, Kothamangalam, Kerala, India e-mail: bobincherian@mace.ac.in C. Rajendran Department of Management Studies, Indian Institute of Technology Madras, Chennai, India e-mail: craj@iitm.ac.in H. Ziegler School of Business, Economics and Information Systems, University of Passau, Passau, Germany e-mail: hans.ziegler@uni-passau.de © Springer Nature Switzerland AG 2021 S. Srinivas et al. (eds.), Supply Chain Management in Manufacturing and Service Systems, International Series in Operations Research & Management Science 304, https://doi.org/10.1007/978-3-030-69265-0_6 165 166 B. C. Jos et al. In the present scenario, customers being the focus of the organization, retaining them becomes the primary objective of any organization. Retaining customers can be achieved by meeting and sometimes exceeding their expectations. One of the main reasons for the loss of customer goodwill is the non-conformance to due dates (Unlu and Mason 2010). Many times, either due to the non-availability of enough capacity to process all the jobs or the non-availability of the resource for manufacturing, many manufacturing firms outsource some of their jobs or orders to potential subcontractors located geographically away from the manufacturer. Transportation delay and the associated transportation costs are additional aspects that need to be considered while preparing the production schedule of the jobs when outsourcing is chosen as an option (Qi 2008). In order to reduce the tardiness of jobs, subcontractors usually increase the number of shipments from their locations. Energy and environmental factors need to be considered in many such cases. Most of the manufacturing firms focus on minimizing the cost of production, sometimes by ignoring the resultant carbon emissions. Because of the growing concerns in climate change, manufacturers are now forced to reduce carbon emissions. Hence this study also focuses on minimizing carbon emissions related to the transportation schedules while preparing the operations plan in a two-stage supply chain with outsourcing options. 2 Literature Review We have done an extensive survey on the literature which focus on scheduling of jobs with the consideration of outsourcing options, transportation time, selection of subcontractors, and reduction in carbon emissions. Our study integrates all these decisions which have been considered separately in the literature. The studies of Cai et al. (2005) and Qi (2008) considered transportation time for the movement of jobs while preparing the production schedule. Cai et al. (2005) considered a scheduling problem involving third-party machines having specific time slots, with additional transportation time for the movement of jobs back to the manufacturer. Qi (2008) considered production scheduling vis-a-vis coordination of transport when the production or other services are outsourced, and proposed dynamic programming algorithms. However, Qi (2008) assumed the presence of only a single manufacturing facility located in-house and a single subcontractor. Cao et al. (2005) were probably the first to study a parallel machine scheduling problem associated with the option of machine selection from a set of potential machines, each having a fixed hiring cost. They modeled their problem to minimize the sum of total tardiness cost and the machine holding cost. Cao et al. (2005) studied the problem involving subcontractors (mostly treated as parallel machines, as for the allocation and scheduling of jobs) to fulfill customer interests in time. RuizTorres et al. (2008) observed that many organizations might not have any in-house production facility; rather they would outsource the manufacturing operations. The parent organization or central manufacturing firm controls the assignment of jobs to An Integrated Problem of Production Scheduling and Transportation in a Two-. . . 167 these parallel identical production resources/outsourced centers with its own design and planning functions. The authors quoted Nike as an example. They proposed heuristics to solve this problem and conducted computational experiments. Alidaee and Li (2014) studied a similar problem with an option of choosing a subset of subcontractors from a set of subcontractors, and to schedule the operations in all the chosen subcontractors to minimize the sum of costs involving the cost of holding machines, cost of total machine usage, and total tardiness costs. They considered identical processing times and a common due date for all jobs, and proposed dynamic programming algorithms. Mokhtari et al. (2012) presented a job scheduling problem with outsourcing options and proposed a MILP model to minimize the cost of mean weighted flow time and cost of processing the outsourced orders. However, their model did not consider the transportation of finished products. The study by Wang and Alidaee (2018) considered unrelated parallel machines and scheduling of jobs with the objective of minimizing the total workload. Since the objective was not correlated with minimizing tardiness, the transportation concerns were not significant, and hence they presented mathematical models and heuristics by ignoring the logistics schedule. Research on the reduction in carbon emissions has attained great attention recently because of the growing concerns on climate change. In the report of the United Nation’s Inter-Governmental Panel on Climate Change (UN IIPC), Kahn Ribeiro et al. (2007) states that, by 2030, energy use and carbon emissions from transport are predicted to be 80 percent higher than the current levels. The transportation sector is responsible for approximately 23% of the global energy related carbon dioxide emissions (Kahn Ribeiro et al. 2007). Logistics are expected to make a large contribution to the drastic reductions in CO2 emissions that will be required by 2050 to contain the global temperature increase within 2 ◦ C by 2100 (McKinnon 2010). Ritchie and Roser (2017) have consolidated various studies related to CO2 emissions. The statistics presented in their studies on global CO2 emissions show the significance of reducing such emissions to prevent further increases in carbon emissions. Because of the growing concerns on global warming and climate changes, many countries have implemented a ‘carbon emissions policy’ as mandatory. From Fig. 1, it is evident that the transport sector is the second major contributor to CO2 emissions, following the ‘electricity and heat’ sector. Hence, manufacturers now include ‘reduction in carbon emissions’ as a key criterion while preparing the operations plan. In our study, we also consider it as one of the key criteria while preparing the production schedule. Few of the recent studies in scheduling problems related to ‘reduction in carbon emissions’ are given in this section. The reduction of potential carbon emission in an integrated scheduling problem of ‘picking up and delivering customers to airport’ was studied by Yu et al. (2016). Their study illustrated that integrated scheduling could reduce carbon emissions to a significant level. The problem of vehicle routing and scheduling in a similar environment was studied by Yang and Tang (2013) with the objective of minimizing carbon emissions. They presented a nonlinear 0–1 integer programming model. They assumed that the carbon emission (E) of a vehicle 168 B. C. Jos et al. Global CO2 Emissions by Sector 16 12 Electricity/Heat billion tons 14 Transportaon Manufacturing/Const rucon Building 10 8 Land-Use Change and Forestry Industrial Processes 6 Bunker Fuels 4 Other Fuel Combuson Fugive Emissions 2 0 1990 1995 2000 2005 2010 2015 Source: CAIT Climate Data Explorer Fig. 1 Global CO2 emissions by sector is directly related to the fuel consumption (F) through the relation E = r × F, where r is the carbon emission index parameter. Fang et al. (2011) studied a scheduling problem in manufacturing for power consumption and carbon footprint reduction. They assumed that the operation speed could be changed to factor the peak load and energy consumption. Wang et al. (2019) studied an integrated single machine scheduling and vehicle routing problem for minimizing the total carbon emissions, which allows switching on-off the machine in-between the operations. They proposed a mathematical model and a multi objective tabu search algorithm to minimize the total carbon emissions. In the literature, the research attempts on job scheduling with the selection of outsourced centers have either assumed a common due date for jobs or have not considered the transportation concerns (related to transportation time, transportation cost associated with outsourced facility and resultant carbon emissions), while optimizing the specific objective function. However, those researches that considered transportation concerns assumed a pre-set number of outsourced facilities with identical job processing times. As seen from the review, none of the works have incorporated the transportation parameters concerning the outsourced facilities, while studying the problem of selecting outsourced facilities from the available facilities. We believe that in a supply chain system, the transportation decisions are significant while preparing the production schedule of orders with the objective of minimizing total costs, which includes tardiness costs (Qi 2008). An Integrated Problem of Production Scheduling and Transportation in a Two-. . . 169 Based on the research gap identified, we study an integrated problem of logistics and scheduling of jobs in a supply chain with subcontracting decisions and the associated transportation issues. We present a new mixed-integer linear programming (MILP) model to solve the problem under study with the objective of reducing the total supply chain cost, as well as limiting the total carbon emissions to a prescribed level. Our MILP model is flexible enough to handle scenarios where the central manufacturing firm may or may not have in-house production facilities. Hence the model can also be implemented into a virtual enterprise system as described by Ruiz-Torres et al. (2008). The main contribution of the study is the proposed solution methodology that integrates three different issues, (1) selection of subcontractors, (2) scheduling of jobs, and (3) scheduling of logistics, which have not been considered together in the literature so far. In addition, this study considers ‘carbon emission aspects’ associated with the shipment of finished jobs that is necessary for making the supply chain eco-friendly. From the viewpoint of a supply chain, the external manufacturing facility is an outsourced center of production. From the viewpoint of scheduling theory, these outsourced centers can be referred to as machines in general. Hence in this chapter, the terms ‘manufacturing facility’ (referring to both outsourced and in-house) and ‘machine’ are used interchangeably within the context of our study. In this chapter, the set of parallel in-house manufacturing facilities is referred to as a central manufacturing facility or parent manufacturing facility. 3 Problem Description The integrated logistics and job scheduling problem in a supply chain with nonidentical parallel manufacturing facilities is described as follows. At the beginning of the planning period, the manufacturer (also called central organization) receives a set of orders or jobs J = {1, 2, 3, . . . , N} from its customers. It is assumed that all jobs are available for processing at the beginning of the planning period and need to be processed either by in-house or outsourced facilities of the organization. The central manufacturing center is equipped with Min parallel machines, which we refer to as in-house facilities. In order to minimize the tardiness of jobs (to make an early delivery), they can outsource some of the jobs to subcontractors, each equipped with a single machining facility. We consider them as the outsourced manufacturing facilities of the central organization. Machining facilities of subcontractors need not be identical and are located geographically away from the central organization. Hence, outsourcing will incur transportation delay and transportation costs, in addition to the fixed cost (hiring cost) of outsourcing the facility and the associated variable cost of production. There are Mout potential subcontractors willing to process the jobs as per the design and schedule given by the central organization. We consider the availability of Min in-house machines and the Mout subcontractors’ machines (with each subcontractor having a single machine as the manufacturing facility) 170 B. C. Jos et al. Pool of jobs 1 Given Machining facilies in Parallel A sample set of jobs assigned to each machine In-house machines 2 {12- 1-7} 3 4 {10-3-11} 5 {8-2} 6 7 8 Outsourced Centers 9 10 {4-6-9} 11 12 {φ} 13 14 {5-14} 15 {13-15} Fig. 2 Supply chain system with parallel manufacturing facilities both in-house and outsourced together as a set of available non-identical manufacturing facilities P = {1, 2, 3, . . . , M} working in parallel, where n{P} = M = Min + Mout (see Fig. 2). Some other assumptions with respect to the problem under study are as follows. Each manufacturing facility (both in-house and outsourced) is capable of processing any job, but at most one job at a time. Manufacturing facilities are continuously available, and pre-emption is not allowed. There is no idle time between the processing of jobs on a manufacturing facility, other than the necessary sequence-dependent setup time between the processing of two consecutive jobs. The processing time of an order/job and its setup time depend on the manufacturing facility to which the job is assigned. The resources necessary for processing the jobs are available with the subcontractors, and jobs can be started at any facility without any delay. An Integrated Problem of Production Scheduling and Transportation in a Two-. . . 171 All the jobs manufactured at an outsourced center need to be transported back to the central facility after completion. The number of jobs to be transported in a batch from the outsourced facility to the central facility may be less than or equal to the finite capacity of the transportation vehicle. This decision on how many jobs to be transported per trip depends upon the due date of jobs and also the carbon emission level. This aspect of the problem is unique and significant in our study, and is elaborated further. For each job i ∈ J, the processing time on manufacturing facility m (Pi, m ), the due date (Di ), and the tardiness penalty per unit time (Wi ) are known at the beginning of the planning period. Machine- and sequence-dependent setup times on each manufacturing facility are also known in advance. In this chapter, we hereafter refer to them as setup times. (m ∈ P), the machine ∗ For each manufacturing facility (i.e., fixed subcontracting usage cost per unit time Cm , the fixed hiring cost Cm cost), the transportation time required to ship the outsourced jobs to the central manufacturing facility (Tm ) and the associated cost per trip for transportation (Cm ) are also known in advance. There is a fixed amount of carbon emissions per shipment (Em ) while transporting the finished jobs back to the central facility from the outsourced facility m. Em is proportional to fuel consumption (Yu et al. 2016), and we assume Em is, in effect, proportional to the transportation time taken by the trucks. Hence it is calculated as Em = r × Tm , where r is the carbon emission index parameter. The trucks available at facility m are assumed to be homogeneous and unlimited, and each has a finite capacity, defined in terms of Bm , the number of jobs. Transportation time, transportation cost, fixed hiring cost, and carbon emissions per shipment are assumed to be zero for in-house facilities. Note that ∗ is a constant for a given manufacturing the machine usage cost per unit time Cm facility, and is independent of the jobs. Hence the total machine usage cost of each manufacturing facility is proportional to the sum of the corresponding setup times and the processing times of jobs assigned to that facility. In other words, the total machine usage cost is given by the product of the machine usage cost per unit time and the sum of appropriate setup times and processing times of all jobs assigned to that manufacturing facility. We also assume that a manufacturing facility with better process capability can process the jobs in lesser time but are associated with a higher machine usage cost (both cost per unit time and fixed hiring cost) than a machine with lesser processing capability. However, machine usage cost per unit time of in-house machines is assumed to be less in comparison to all outsourced facilities, irrespective of the processing capability. The decisions to be taken are: (i) to determine the manufacturing facilities (both in-house and outsourced) to be chosen and engaged for processing; (ii) to determine the subset of jobs to be processed on the selected manufacturing facilities (both in-house and outsourced); (iii) to prepare the production schedule of jobs on each manufacturing facility; and (iv) to prepare the transportation schedule of completed jobs at each outsourced manufacturing facility. We present this problem with the objective of minimizing the sum of fixed cost of hiring manufacturing facilities, their usage cost (variable cost), tardiness penalties (variable cost), and transportation costs incurred while shipping finished jobs from 172 B. C. Jos et al. outsourced centers (related to fixed cost per trip from each outsourced center to the central manufacturing firm). 4 Development of the Proposed Mixed Integer Linear Programming Model A Mixed Integer Linear Programming (MILP) model is proposed to solve the supply chain problem under study, and is described in this section. Parameters: Pi, m Di Wi Tm Cm Cm ∗ Cm S0, i, m Si ,i,m i, i Bm Em EUB EL Fm G Processing time for job i on manufacturing facility m. Due date for job i. Tardiness penalty per unit time for job i. Transportation time required to transport a job from manufacturing facility m to the central manufacturing firm /* Tm = 0 if m ∈Min (the set of in-house machines) */. Fixed transportation cost required to transport a job from manufacturing facility m to the central manufacturing firm /* Cm is non-decreasing with respect to Tm ; Cm = 0 if m ∈ Min */. = 0 if m Fixed hiring cost associated with manufacturing facility m /* Cm ∈Min */. Machine usage cost per unit time for manufacturing facility m. Setup time for job i on manufacturing facility m, where job i is the first job to be processed on facility m /* fictitious job i = 0 represents the beginning of processing on a facility and S0, i, m is a positive integer */. Setup time for job i processed immediately after job i on manufacturing facility m. Indices for jobs, i ∈ J and i ∈ J. Capacity restriction in a truck or Maximum batch size (given in terms of the maximum number of jobs transported per trip) from a given facility m /* it assumes that the capacity of vehicle available at an outsourced facility can be different from the capacity of vehicle available at a different outsourced facility */. Carbon emission associated with each trip from manufacturing facility m to the central facility. Upper bound on the total carbon emission, EUB = N × maxm (Em ) Desired limit for the total carbon emissions set by the decision makers. EL = fr × EUB where fr denotes a fraction (0 < fr ≤ 1) Capacity of manufacturing facility m (in terms of the number of jobs that can be processed during the scheduling period). A sufficiently large positive integer, where N N G≥ N i=1 maxm pi,m + i=1 maxi ,m si ,i,m + i=1 maxm s.0,i,m An Integrated Problem of Production Scheduling and Transportation in a Two-. . . 173 Note: n{J} = N and n{P} = M, where J is the set of jobs and P is the set of available manufacturing facilities (both in-house and outsourced facilities included). Binary Variables: Δm Φ i, m γi ,i,m γN+m,i γ i, N + M + m θ i, b, m ∂ b, m a binary variable that takes the value 1 if manufacturing facility m is selected for processing jobs; 0 otherwise. a binary variable that takes the value 1 if manufacturing facility m is selected and job i is assigned to facility m; 0 otherwise. a binary variable that takes the value 1 if job i is processed immediately after job i on the manufacturing facility m to which both jobs are assigned; 0 otherwise. a binary variable that takes the value 1 if job i is processed as the first job on the selected manufacturing facility m; 0 otherwise. a binary variable that takes the value 1 if job i is processed as the last job on a selected manufacturing facility m; 0 otherwise. a binary variable that takes the value 1 if order i is assigned to facility m and transported to the central facility in batch shipment b; 0 otherwise. a binary variable that takes the value 1 if any job is assigned to facility m and transported to the central facility in batch shipment b; 0 otherwise. Decision variables: si oi ti t i, m ∗ tb,m t i μm τi start time of job i. occupancy time of job i /* the sum of the appropriate machine- and sequencedependent setup time and the processing time of job i on its assigned machine at a manufacturing facility */. completion time of job i. time at which job i is shipped from machine m (equals its completion time ti if processed in-house; t i, m ≥ ti for outsourced facility m). completion time of the last job assigned to batch b of machine m /* time of shipment of batch b from facility m */. time at which job i is delivered at the central facility. total machine usage time of a manufacturing facility m. tardiness of job i. Objective function: Min Z = N Wi τi + i=1 M Cm m + m=1 N M Cm ∂b,m + b=1 m=1 M ∗ Cm μm (1) m=1 subject to the following: M m=1 i,m = 1 ∀i ∈ J (2) 174 B. C. Jos et al. N γi ,i,m + γN+m,i = i,m ∀i ∈ J, ∀m ∈ P (3) i =1, i =i N γi,i ,m + γ i,N+M+m = i,m ∀i ∈ J, ∀m ∈ P (4) i =1, i =i N γN+m,i ≤ m ∀m ∈ P (5) γi,N+M+m ≤ m ∀m ∈ P (6) i=1 N i=1 N i,m ≤ N × m ∀m ∈ P (7) i=1 N i,m ≥ m ∀m ∈ P (8) i=1 N i,m ≤ N i=1 N N γN+m, i ∀m ∈ P (9) ∀m ∈ P (10) i=1 i,m ≤ N i=1 N γi,N+M+m i=1 si ≥ ti − G 1 − M ∀i ∈ J, i ∈ J and i = i γi ,i,m (11) m=1 si ≤ G 1 − γN+m, i oi = N M ∀i ∈ J, ∀m ∈ P γi ,i,m Si ,i,m + i ,i =i m M γN+m, i S0,i,m + m M Pi,m i,m (12) ∀i ∈ J m (13) ti = si + oi M ti,m ≥ ti ∀i ∈ J ∀i = J (14) (15) m=1 N b=1 θi,b,m = i,m ∀i = J, ∀m ∈ P (16) An Integrated Problem of Production Scheduling and Transportation in a Two-. . . ∗ tb,m ≤ G × ∂b,m ∂b,m ≥ ∂(b+1),m N 175 for b = 1, .., N, ∀m ∈ P (17) for b = 1, .., N − 1, ∀m ∈ P (18) ∂b,m ≤ N × m ∀m ∈ P (19) b=1 m ≤ N ∀m ∈ P ∂b,m (20) b=1 ∂b,m ≤ N θi,b,m ∀m ∈ P , ∀b = 1, .., N (21) ∀m ∈ P , ∀b = 1, .., N (22) i=1 N θi,b,m ≤ N × ∂b,m i=1 ∗ ≥ tb,m − G 1 − θi,b,m ti,m ∀i = J, ∀b = 1, .., N, m ∈ P (23) ∗ ti,m ≤ tb,m + G 1 − θi,b,m ∀i = J, ∀b = 1, .., N, m ∈ P (24) ti,m ≤ G × i,m ti = M ti,m + m=1 M ∀i = J, ∀m ∈ P ∗ tb,m ≤ G × ∂b,m τi ≥ t i − Di μm ≥ ti,m μm ≤ G × m i=1 θi,b,m ≤ Bm ∀i = J (26) ∀b = 2, .., N, ∀m ∈ P (27) Tm i,m m=1 ∗ ∗ ≥ t(b−1),m − G 1 − ∂b,m tb,m N (25) ∀b = 2, .., N, ∀m ∈ P (28) ∀i ∈ J (29) ∀i ∈ J, ∀m ∈ P (30) ∀m ∈ P (31) ∀m ∈ P , ∀b = 1, .., N (32) 176 B. C. Jos et al. N i,m ≤ Fm ∀m ∈ P (33) Em ∂b,m ≤ E L ∀m ∈ P , ∀b = 1, .., N (34) i=1 N M b=1 m=1 i,m ∈ {0, 1} ; γi ,i,m ∈ {0, 1} ; γN+m, i ∈ {0, 1} ; γ i,N+M+m ∈ {0, 1} ; m ∈ {0, 1} ; ∂b,m ∈ {0, 1} ; θi,b,m ∈ {0, 1} ; ∀m ∈ P , ∀i ∈ J, ∀i ∈ J and i = i; and all other variables are ≥ 0.Fm and Bm are integers. (35) Equation (1) minimizes the total costs linked with the tardiness of jobs, the total fixed hiring costs of the manufacturing facilities selected, the total transportation costs incurred while shipping the jobs processed at outsourced facilities back to the central organization in batches, and the total machine usage costs of the facilities selected. Constraint (2) reflects that a job can be assigned to at most only one manufacturing facility. The feasibility of the sequence is maintained by Eqs. (3) and (4). Constraints (5) and (6) guarantee that each manufacturing facility, if selected, can have at most one start job as well as one end job. These jobs are assumed as fictitious jobs and are named as (N + j) and (N + J + j) respectively. Constraints (7) and (8) warrant that if at least one job is allocated to a manufacturing facility, then the facility is active; and that a manufacturing facility is inactive if no job is allocated to it. Constraints (9) and (10) ensure the allocation of fictitious jobs to a manufacturing facility if any of the jobs are allocated to that facility. Constraint (11) guarantees that the start time of any job is greater than or equal to the completion time of its predecessor. Constraint (12) assigns the start time of the first job on each manufacturing facility to zero. Equations (13) and (14) compute the completion time of each job as the sum of the respective start time, setup time, and processing time. The shipment of any job from a manufacturing facility is possible only after the completion of that job at the respective facility. This condition is ensured by constraint (15). Constraint (16) states that shipment of a job in any batch from a manufacturing facility is possible only if the job is processed in that facility. Constraint (17) ensures that completion time of the last job assigned to any batch shipment from a facility can be considered as a positive integer, only if at least one job is assigned to that batch shipment from the respective facility. ‘Compression constraint’ (18) ensures that a positive batch-quantity shipment cannot follow a zero batch-quantity shipment. Constraints (19) and (20) ensure that the batch shipment from a manufacturing facility is possible only if that facility is active. Constraints (21) and (22) ensure that a batch shipment from a manufacturing facility is active only if jobs are assigned to that batch from the respective facility. Constraints (23) and (24) state that the shipment time of a job from a manufacturing facility is equal to its corresponding batch’s shipment time from the respective facility. Constraint (25) ensures that the shipment time of any job from a facility is a positive integer only if that job is allocated to that facility. Constraint (26) states that the arrival time An Integrated Problem of Production Scheduling and Transportation in a Two-. . . 177 of a job at the central manufacturing facility is the sum of shipment time from the manufacturing facility and the associated transportation time. Constraint (27) and (28) ensure that the completion times of different batches from a facility are in the non-decreasing order of batch numbers. Expression (29) determines the tardiness of all jobs. Expressions (30) and (31) define the total machine usage time of each manufacturing facility selected. Constraints (32) and (33) maintain the batch capacity and manufacturing capacity restrictions, respectively. Constraint (34) is used to keep the total carbon emission from transportation within the desired upper limit. Finally, Expression (35) defines all the binary variables and continuous variables. 5 A Numerical Example In this section, we present a numerical example of the problem under study and its solution using the proposed solution technique. Consider a supply chain system with number of jobs (N) = 8 and number of machines (M) = 2, where in-house machines Min = 1 and outsourced centers Mout = 1. This example is meant only for the illustration purpose, and hence units of the input parameters are not specified. Processing time of each job i on machine j (Pi, j ), due date associated with job i (Di ), and tardiness cost/unit time (Wi ) are given in Table 1. Sequence-dependent setup times (Si ,i,m ) on machines m = 1 and 2 are given in Tables 2 and 3, respectively. Other input parameters are given in Table 4. The maximum capacity of a truck from facility m (Bm ) is assumed as 3 jobs and is same for all m, and the maximum production capacity (Fm ) at a manufacturing facility is 6 jobs and is same for all m. The carbon emission limit is calculated as 80 using the equation EL = fr × EUB where fr is assumed to be 0.5. The values of the different parameters are generated randomly, and is described in section 6. The given problem instance is solved using the proposed MILP model and using IBM CPLEX optimization studio 12.7.1. The objective function value of the optimal Table 1 Input parameters: processing time, due date, and tardiness cost Job no. (i) 1 2 3 4 5 6 7 8 Processing times on facilities (Pi, m ) Pi, 1 Pi, 2 11 12 6 6 9 10 10 11 9 10 3 3 6 6 9 10 Due date (Di ) 38 39 36 37 34 34 37 41 Tardiness Cost (Wi ) 12 11 15 12 10 10 12 12 178 Table 2 Sequence-dependent setup time (Si ,i,m ) on facility m=1 Table 3 Sequence-dependent setup time (Si ,i,m ) on facility m=2 Table 4 Other input parameters B. C. Jos et al. Preceding job (i ) 0 1 2 3 4 5 6 7 8 Preceding job (i ) 0 1 2 3 4 5 6 7 8 Succeeding job (i) 1 2 3 4 5 4 2 2 6 4 0 3 5 5 5 4 0 2 6 5 4 5 0 5 4 2 2 4 0 6 3 5 4 5 0 4 5 2 6 3 4 5 3 2 3 3 2 3 3 5 6 2 5 6 2 5 3 0 2 5 7 3 2 2 5 6 4 2 0 3 8 2 3 4 2 4 6 4 2 0 Succeeding job (i) 1 2 3 4 5 4 2 2 6 4 0 3 5 5 5 4 0 2 6 5 4 5 0 5 4 2 2 4 0 6 3 5 4 5 0 4 5 2 6 3 4 5 3 2 3 3 2 3 3 5 6 2 5 6 2 5 3 0 2 5 7 3 2 2 5 6 4 2 0 3 8 2 3 4 2 4 6 4 2 0 Manufacturing facility (m) Transportation time (Tm ) Transportation cost (Cm ) Fixed cost of hiring Cm ∗ Machining cost/unit time Cm Carbon emission/trip (Em ) 1 0 0 0 2 0 2 20 60 240 8 20 solution is obtained as 1019 in a computational time of 54.45 sec. Jobs 2 and 8 are shipped from manufacturing facility 2 in a batch. The machine usage time of facilities 1 and 2 are 63 and 20, respectively. In the optimum schedule, tardiness corresponding to jobs 1, 2, and 5 are found to be 11, 1, and 29, respectively. All other jobs have met their respective due dates. Allocation and the schedule of jobs on each machine, when selected, are shown in the following Gantt chart (Fig. 3). 6 Computational Experimentation We designed our experiments to generate problem sets randomly as suggested by earlier researchers, especially by Potts and Van Wassenhove (1985), Liaw et al. (2003), Chen (2009) and Oğuz et al. (2010). In this section, we first explain how An Integrated Problem of Production Scheduling and Transportation in a Two-. . . 179 Indicates the setup process of a job Facilities\Machines i 2 8 1 3 5 Indicates the processing of Job i 2 6 10 15 7 20 4 25 30 1 35 40 45 5 50 55 60 Time Fig. 3 Gantt chart representation of the optimal solution to generate the problem sets, and then demonstrate the results of the computational analysis obtained using the proposed solution technique. The values of the different parameters for the integrated logistics and job scheduling problem in a supply chain with non-identical parallel manufacturing facilities are generated as given below. For each order i ∈ J, the processing time Pi, m on each facility m ∈ P is an integer generated uniformly in the interval [1, 20]. The sequence-dependent setup times Si ,i,m and initial setup times S0, i, m are also integers generated uniformly in the interval [1, 10]. Triangular inequality is maintained while generating setup times. We assume that Pi, m depends on the machining facility at a manufacturing center and a machine performance factor fm determines the relative performance of the facilities, thereby the processing time of a job. fm also influences the sequence-dependent setup times at each manufacturing facility (Chen 2009). We consider the machine performance factor fm as a uniform random number in the interval [2/3, 3/2]. The following equations are used to incorporate the machine performance factor while generating the processing times and setup times to ensure that the parameter values are within the interval specified 3 earlier: Pi, m = fm × Pi, v and Si,j,m = fm × Si ,i,v where Pi,v = U 2 × 1, 32 × 20 and Si ,i,v = U 32 × 1, 32 × 10 . Here, v denotes a fictitious machine, Pi, v denotes the processing time of job i on machine v, and Si ,i,v denotes the appropriate setup times on the machine v. The due date (Di ) is set to maxi ,m Si ,i,m + max slack, maxm Pi,m + avgm (Tm ) , where slack is generated uniformly in the interval Pa [1 − Π − R/2 M 2 1 − Π + R/2] where Pa = N i=1 m=1 Pi,m /M (Liaw et al. 2003). The job due- 180 B. C. Jos et al. date is designed to avoid tardiness if a given job is assigned as the first job on any facility without the consideration of transportation time. The parameter Π is called tardiness factor and R is called due-date range, and both parameters are chosen from the values 0.3 and 0.7 (Potts and Van Wassenhove 1985). For each facility m available for the given jobs (i.e., m ∈ P), processing the ∗ , the fixed hiring cost of a facility C , machine usage cost per unit time Cm m the transportation time necessary for jobs to ship from the outsourced center to the central organization (Tm ) and the associated cost for transportation (Cm ) are generated as described. We also assume that manufacturing facilities with advanced processing techniques can process the jobs in lesser time, but with a higher per unit machining cost and holding cost. However, machine usage costs of in-house facilities are assumed to be the least, irrespective of the machine performance. The transportation time, transportation cost, and the fixed holding cost are assumed to be zero for the in-house machines. The usage cost of in-house machine ∗ ) is an integer generated uniformly in the interval [1, 5]. The parameter values (Cin associated with the outsourced facilities are integers and are calculated as follows. Transportation time (Tm ) to ship the finished jobs to the central organization is generated as a uniform random number in the interval [1, 20], corresponding transportation cost (Cm ) is calculated as Ω × Tm and the usage cost of outsourced ∗ ) is calculated as C ∗ × ∗ . For a specific problem instance, Ω and facility (Cout in fm Ω ∗ are constants and are generated randomly in the interval [1, 3] and [2, 5], respectively. This ensures the production cost to be always higher at the outsourced centers than the in-house center, and also in the increasing order of the machine ∗ is a constant for performance factor. Note that the usage cost per unit time Cm a facility and is independent of the jobs. Fixed cost of outsourced facilities is = 30 × C ∗ (30 is based on the sum of the upper bounds for calculated as Cm m processing time and setup time). Tardiness penalty for job i (Wi ) is calculated as ∗ where i is a uniform random number in the interval [1, 2] i × maxm Cm associated with each job i. Tardiness penalty increases with the higher value of i . Hence more jobs will be assigned to outsourced machines, and also more outsourced facilities will be selected. Carbon emissions per shipment from facility m (Em ) is calculated as r × Tm × Vavg , where r is the carbon index parameter (the average amount of carbon emissions per unit distance) for a truck and Vavg is the average speed of a truck (Yu et al. 2016). To analyze the effectiveness of the proposed solution technique, problem instances are generated with three different job sizes: J∈ {8, 10 and 12}. In all these instances, we consider parallel manufacturing facilities; M = 3 with two levels of in-house facilities (0 and 1). Three problem instances are generated for each combination of tardiness factor (Π) and due date range (R) chosen from the values 0.3 and 0.7. Hence, by considering three sets of jobs (J= 8, 10 and 12) in conjunction with the two levels of Π and R (each: 0.3 and 0.7), and two levels of in-house facilities (0 and 1), this study generates a total of 72 problem instances. The proposed MILP model is tested using the commercial software CPLEX 12.7.1, and on a system with 64 bit Intel i5 3.2 GHz Processor and 8.0 GB of RAM. A time restriction of 7200 seconds is enforced while executing the MILP model on An Integrated Problem of Production Scheduling and Transportation in a Two-. . . 181 CPLEX. For each problem instance, we observe the objective function value and the execution time required for the solver to converge optimally under the given system environment. If the problem instance has not been converged to optimum within the stipulated time, we observe the % optimality gap reported by the solver. Computational analysis of the proposed MILP model is reported in three dimensions, (i) the average execution time taken to solve the test instances optimally within the given time limit of 7200 seconds, (ii) the number of problem instances solved to optimality within the given time limit in a set of 3 problem instances with similar parameters, and (iii) the average optimality gap (%) reported by the solver during termination. Results are compiled and given in Table 5. An analysis of the experimental results given in this section shows the effectiveness of the proposed MILP model. In small problem instances, the proposed model could give optimal solutions within a short time. When the job size is increased to 12, the model could not converge to an optimum value within the stipulated time Table 5 Computational analysis N×M 8×3 Min 0 1 10 × 3 0 1 12 × 3 0 1 Π ×R 0.3 × 0.3 0.3 × 0.7 0.7 × 0.3 0.7 × 0.7 0.3 × 0.3 0.3 × 0.7 0.7 × 0.3 0.7 × 0.7 0.3 × 0.3 0.3 × 0.7 0.7 × 0.3 0.7 × 0.7 0.3 × 0.3 0.3 × 0.7 0.7 × 0.3 0.7 × 0.7 0.3 × 0.3 0.3 × 0.7 0.7 × 0.3 0.7 × 0.7 0.3 × 0.3 0.3 × 0.7 0.7 × 0.3 0.7 × 0.7 Average time (s)a 182 80 140 176 17 32 145 214 – 5534 5307 5555 – 564 5422 3598 – – – – – – – – Number of instances solved to optimality 3 3 3 3 3 3 3 3 0 1 1 1 0 3 1 1 0 0 0 0 0 0 0 0 Average optimality gap (%) 0 0 0 0 0 0 0 0 26.3 11.9 29.3 20.4 30.5 0 12.3 15.8 50.9 48.2 64.5 68.8 69.6 61.4 65.4 72.7 Character ‘-’ indicates that none of the problem instances reached optimality within 7200 s for computing average CPU time (i.e., execution time for all instances is above 7200 s) 182 B. C. Jos et al. results in optimality gaps. In the computational analysis, it is observed that the presence of in-house machines assigns more jobs to such machines resulting in a reduction in total transportation cost as well as total carbon emissions. Larger due dates and more jobs can be assigned to in-house machines because jobs assigned to in-house machines can meet the due dates easily due to the absence of additional transportation time. A lower value of the ‘carbon emission limit’ will also assign more jobs to the in-house machines. However, such an allocation will account for a greater number of tardy jobs. The due dates depend on the processing times, setup times, tardiness factor, range factor, and transportation times generated randomly. Hence, in our experimentation, we observe that a variation in a single parameter alone cannot make any significant difference in computational time. The more the tardiness penalty associated with jobs, the more facilities will be selected, and more jobs will be assigned to outsourcing facilities. 7 Summary Scheduling theory assumes great significance in the supply chain planning activities of an organization involved with in-house and outsourced manufacturing facilities. In this chapter, we have studied an integrated problem of logistics and scheduling of jobs with options of outsourcing facilities and additional consideration of transportation decisions related to transportation cost, transportation time, and carbon emissions. The present study has integrated three different issues in a supply chain: the selection of subcontractors, the scheduling of jobs, and the scheduling of logistics. The main contribution of the study is the proposed solution methodology that integrates these three different issues which have not been considered together in the literature so far. In addition, this study considers ‘carbon emission aspects’ associated with the shipment of finished jobs that is necessary for making the supply chain eco-friendly. In this supply chain system, we need to choose a set of subcontractors (also called outsourced manufacturing facilities) from the available subcontractors to process a part of its orders/jobs, in addition to the manufacturing facilities located in-house, to maximize the supply chain profitability while considering carbon emissions. All the completed jobs at the outsourced centers need to be transported back to the central manufacturing firm. The orders/jobs to be processed are having different processing times and due dates. We assume that the manufacturing facilities, both outsourced and located in-house, can be non-identical and are working in parallel. We have proposed a MILP model capable of giving optimum solutions for small problem instances (up to job size 10) within a reasonable time. Precedence-based binary variables are used in the proposed MILP formulation to denote the sequence of jobs in a manufacturing facility. A numerical example and its solution have been presented to describe the problem scenario. Computational analysis with sample problem instances is conducted to test the robustness of the proposed solution An Integrated Problem of Production Scheduling and Transportation in a Two-. . . 183 methodology. The computational time required to solve the problem instances to optimality is found to be more for large-sized problem instances. The proposed model is useful for organizations with make-to-order policies. The previous scheduling models mainly optimize production-related costs. They ignore the possibility of reducing the delay in product delivery (tardiness of jobs). The proposed model not only optimizes the production-related costs, but also helps to reduce the tardiness of jobs and determine the related transportation decisions and emission levels in a supply chain. Our MILP model is flexible enough to handle scenarios where the central manufacturing firm may or may not have inhouse production facilities. Hence the model can also be implemented into a virtual enterprise system as described by Ruiz-Torres et al. (2008). Future studies can include the development of heuristics and meta-heuristics for this integrated problem of logistics and job scheduling with outsourcing options. Further, the research can also include the development of solution methodologies to this supply chain problem with the additional consideration of stepwise transportation cost. The optimal solutions obtained for small-sized problems and the lower bounds obtained for large-sized problems using the proposed solution methodology can be used in the future to evaluate the performance of heuristics, which might be developed for similar problems under study. References Alidaee, B., & Li, H. (2014). Parallel machine selection and job scheduling to minimize sum of machine holding cost, total machine time costs, and total tardiness costs. IEEE Transactions on Automation Science and Engineering, 11(1), 294–301. https://doi.org/10.1109/ TASE.2013.2247757. Azadian, F., Murat, A., & Babu, R. (2015). Integrated production and logistics planning: Contract manufacturing and choice of air/surface transportation. European Journal of Operational Research, 247(1), 113–123. https://doi.org/10.1016/j.ejor.2015.06.004. Cai, X., Lee, C. Y., & Vairaktarakis, G. L. (2005). Optimization of processing and delivery decisions involving third-party machines. Nonlinear Analysis, Theory, Methods and Applications, 63(5–7), 2269–2278. https://doi.org/10.1016/j.na.2005.03.067. Cao, D., Chen, M., & Wan, G. (2005). Parallel machine selection and job scheduling to minimize machine cost and job tardiness. Computers & Operations Research, 32(8), 1995–2012. https:// doi.org/10.1016/j.cor.2004.01.001. Chen, J. F. (2009). Scheduling on unrelated parallel machines with sequence- and machinedependent setup times and due-date constraints. International Journal of Advanced Manufacturing Technology, 44(11–12), 1204–1212. https://doi.org/10.1007/s00170-008-1917-3. Chopra, S., Meindl, P., & Kalra, D. V. (2016). Supply chain management: Strategy, planning, and operation (6th ed.). Harlow: Pearson. Fang, K., Uhan, N., Zhao, F., & Sutherland, J. W. (2011). A new approach to scheduling in manufacturing for power consumption and carbon footprint reduction. Journal of Manufacturing Systems, 30(4), 234–240. https://doi.org/10.1016/j.jmsy.2011.08.004. Kahn Ribeiro, S., Kobayashi, S., Beuthe, M., Gasca, J., Greene, D., Lee, D. S., Y. Muromachi, P. J. Newton, S. Plotkin, D. Sperling, R. & Wit, R. (2007). Transport and its infrastructure. In: Netz, B., Davidson, O. R., Bosch, P. R., Dave, R., & Meyer, L. A. (eds) Climate change 2007: mitigation. Cambridge University Press, Cambridge, UK, pp 324–385. 184 B. C. Jos et al. Liaw, C.-F., Lin, Y.-K., Cheng, C.-Y., & Chen, M. (2003). Scheduling unrelated parallel machines to minimize total weighted tardiness. Computers and Operations Research, 30, 1777–1789. https://doi.org/10.1016/S0305-0548(02)00105-3. McKinnon, A. (2010). The role of government in promoting green logistics. In A. McKinnon, S. Cullinane, M. Browne, & A. Whiteing (Eds.), Green logistics: Improving the environmental sustainability of logistics, Kogan Page, London, 341–360. doi:https://doi.org/10.1080/ 01441647.2010.537101. Mokhtari, H., Nakhai, I., Abadi, K., & Amin-Naseri, M. R. (2012). Production scheduling with outsourcing scenarios: A mixed integer programming and efficient solution procedure. International Journal of Production Research, 50(19), 5372–5395. https://doi.org/10.1080/ 00207543.2011.627687. Oğuz, C., Salman, F. S., & Yalçin, Z. B. (2010). Order acceptance and scheduling decisions in make-to-order systems. International Journal of Production Economics, 125(1), 200–211. https://doi.org/10.1016/j.ijpe.2010.02.002. Pinedo, M. L. (2012). Scheduling: Theory, algorithms, and systems (4th ed.). New York: Springer. Potts, C. N., & Van Wassenhove, L. N. (1985). A branch and bound algorithm for the Total weighted tardiness problem. Operations Research, 33(2), 363–377. Qi, X. (2008). Coordinated logistics scheduling for in-house production and outsourcing. IEEE Transactions on Automation Science and Engineering, 5(1), 188–192. Ritchie, H., & Roser, M. (2017). CO2 and greenhouse gas emissions. Our World in Data. Retrieved June 22, 2020, from https://ourworldindata.org/co2-and-other-greenhouse-gas-emissions Ruiz-Torres, A. J., López, F. J., Ho, J. C., & Wojciechowski, P. J. (2008). Minimizing the average tardiness: The case of outsource machines. International Journal of Production Research, 46(13), 3615–3640. https://doi.org/10.1080/00207540601158799. Unlu, Y., & Mason, S. J. (2010). Evaluation of mixed integer programming formulations for non-preemptive parallel machine scheduling problems. Computers and Industrial Engineering, 58(4), 785–800. https://doi.org/10.1016/j.cie.2010.02.012. Wang, H., & Alidaee, B. (2018). Unrelated parallel machine selection and job scheduling with the objective of minimizing total workload and machine fixed costs. IEEE Transactions on Automation Science and Engineering, 15(4), 1955–1963. https://doi.org/10.1109/TASE.2018.2832440. Wang, J., Yao, S., Sheng, J., & Yang, H. (2019). Minimizing total carbon emissions in an integrated machine scheduling and vehicle routing problem. Journal of Cleaner Production, 229, 1004– 1017. https://doi.org/10.1016/j.jclepro.2019.04.344. Yang, P., & Tang, J. (2013). Minimizing carbon emissions through vehicle routing and scheduling in the shuttle service of picking up and delivering customers to the airport. Acta Automatica Sinica, 39(4), 424–432. https://doi.org/10.1016/S1874-1029(13)60042-7. Yu, Y., Tang, J., Li, J., Sun, W., & Wang, J. (2016). Reducing carbon emission of pickup and delivery using integrated scheduling. Transportation Research Part D, 47, 237–250. https:// doi.org/10.1016/j.trd.2016.05.011. A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for Improving Pharmaceutical Service Alexander Jackson and Sharan Srinivas 1 Introduction Over the past decade, the market has seen increased growth owing to direct-toconsumer (D2C) deliveries; more specifically, e-commerce has grown between 7 and 10 percent in developed markets, and 300% in developing ones (Dayarian et al. 2020). Shopping in e-commerce is primarily focused on speed, flexibility, security, and cost of delivery. Notably, most online shoppers are focused on the speed of delivery while remaining price sensitive. It is estimated that around 50% of U.S adults are more likely to purchase if same-day delivery is offered, and 40% expect the option to be available (Epps and Jordan 2015). However, 88% of consumers prefer free shipping over faster shipping (Deloitte 2018). As consumers balance their demand for cheap and quick delivery, retailers must bear most of the costs of delivery. This has caused retailers to search for cheap alternatives to provide a speedy delivery service (Dayarian et al. 2020). Furthermore, this sort of environment provides a near perfect opportunity for drone applications that can serve the commercial needs of retailers and consumers. Traditional delivery systems that employ trucks are subject to various factors that hurt consumers and limit their effectiveness. Trucks are limited by a region’s A. Jackson () Department of Industrial and Manufacturing Systems Engineering, University of Missouri, Columbia, MO, USA e-mail: atjb47@mail.missouri.edu S. Srinivas Department of Industrial and Manufacturing Systems Engineering, University of Missouri, Columbia, MO, USA Department of Marketing, Trulaske College of Business, University of Missouri, Columbia, MO, USA e-mail: srinivassh@missouri.edu © Springer Nature Switzerland AG 2021 S. Srinivas et al. (eds.), Supply Chain Management in Manufacturing and Service Systems, International Series in Operations Research & Management Science 304, https://doi.org/10.1007/978-3-030-69265-0_7 185 186 Table 1 Summary of drone specifications A. Jackson et al. Drone company Matternet DHL Parcel Zipline Payload (kg) 2 4 1.75 Range (km) 10 65 80 Speed (km/h) 40 130 100 infrastructure as their only way to deliver their products is by road. This increases delivery times as there is often traffic or congestion blocking the roads. Secondly, trucks require a high investment cost to purchase the vehicles as a typical delivery truck costs $50,000 (Hawes 2016). Thirdly, trucks are burdened by high maintenance costs for operation as the average cost of operating a truck is $0.08 cents per mile (Barnes and Langworthy 2004); Furthermore, they require hundreds of dollars in terms of fuel costs. Lastly, trucks also have high labor costs associated with them as each vehicle requires a human operator. UPS drivers are paid up to $25 an hour (Welch 2015), which translates to $52,000 per year. All these costs can be saved by considering alternate delivery methods, especially drones. Drones made by companies such as Matternet are unmanned and batterypowered flying vehicles that use multiple rotors, but drones made by DHL Parcel and Zipline, while unmanned and battery-powered, are fixed winged and designed to be like small planes. Their sizes are typically small but can vary depending on their type and use. Recently, drones have been typically used for surveillance, photography, entertainment, civilian, and military purposes. Innovations in drone hardware, software, and networks have allowed for the continued exploration of drone utilization. Furthermore, the different payloads, range, and speeds of different types of drones currently in use, as shown in Table 1, adapted from Scott and Scott (2017), Edenhofer (2018), Zipline (2020), have allowed these advances to serve market demand. Drones have the potential to lower labor costs, decrease delivery times, and reduce maintenance costs as, in general, they are less expensive to maintain and can operate autonomously (Dorling et al. 2016). Furthermore, drones are not limited to traditional infrastructure such as roads; hence, face less obstacles in their delivery networks. This allows for a speedy and cost-effective way for goods to be delivered. Although commercial delivery with drones has many uses, commercial utilization is not the only service that drones can provide to society. Medication delivery for healthcare can greatly benefit from the implementation of drone technology. Especially during emergencies, drones can deliver vital supplies such as food, water, and medications to disaster areas by flying over damaged infrastructure (Dorling et al. 2016). The timely delivery of medications, blood, and vaccines is critical in healthcare, and some innovative organizations have begun to use drones for healthcare services. Matternet is one such company that uses autonomous drones to deliver medical supplies (Scott and Scott 2017). Although emergencies are a viable application for drones, they do not occur enough to warrant the quick implementation of drones. Pharmacy prescriptions, on the other hand, offer a unique opportunity for possible drone delivery as quick delivery can A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for. . . 187 better serve the community, and 47% of consumers are interested in a drone-based delivery service (ReportLinker 2016). The most significant challenge facing drone distribution networks is the optimization of different drone delivery strategies. Every region has different variables that can affect the strategy implementation for drone deliveries such as infrastructure, urban density, and terrain. However, one strategy utilized by Matternet is the “truck-drone tandem” delivery system. A truck-drone tandem delivery system focuses on the drone and delivery truck working together to deliver their goods. Since the drone has a limited battery life and range, the delivery truck enables the drone to be launched from the truck at locations from where the drones can effectively complete their deliveries. This method allows for long-distance deliveries as it relies on the truck for the initial leg of the delivery, and only allows the drone to take off from the truck for the “last-mile” of the delivery. Furthermore, this method allows for more packages to be delivered farther away from the distribution centers as the truck can hold multiple packages. However, many issues also arise from the tandem method. The initial leg of the trip is still subjugated to all the problems of a traditional delivery method such as poor infrastructure and traffic, high maintenance and labor costs, and high delivery times. Therefore, the tandem method may be optimal in certain scenarios, but other methods need to be considered to have the most effective delivery system with regards to delivery time and cost. A drone delivery method that solely utilizes drones for the delivery of goods can also be simulated. The Drone-Only method bypasses the limitations of the truck entirely; therefore, it can be more optimal in certain regions. A Drone-Only method will eliminate the limitation of infrastructure, long delivery times, and reduce the capital requirements, maintenance, and labor costs associated with trucks. However, a greater number of drones will be required to serve market demand, and the delivery distances will be restricted by the drone’s limited battery life and range. This paper seeks to determine the most optimal drone delivery method for the delivery of medical prescription drugs with regards to delivery time and cost. A discrete-event simulation is utilized for a variety of scenarios, and the different delivery methods are compared. Furthermore, the optimal number of drones required for different areas are determined, and a cost-benefit analysis is performed against traditional delivery trucks. This will provide the information needed for municipalities to determine the validity of a drone integrated delivery system. 2 Literature Review Research has been conducted on multiple fronts for the implementation and application of drone technology for delivery services. Drone routing studies such as Rabta et al. (2018), Ha et al. (2018) focus on developing mathematical models to explore the delivery networks where drones work in tandem with trucks. While they focus on different aspects of the Truck-Tandem strategy, they also build upon research that has not been conducted before. Furthermore, Rabta et al. (2018) focuses on the 188 A. Jackson et al. application of drone delivery in healthcare. The chapter discusses different drone technologies that various companies have developed and also compares them with each other. Furthermore, the chapter also discusses a Truck-Tandem strategy where a truck will deliver to a “drone nest”, from where the drones will perform the lastmile of the delivery. The goal is to minimize the time it takes for the truck and the drone to complete their operation. Next, Dayarian et al. (2020) introduces an innovative technique of having a drone to re-supply a delivery truck while making a delivery. This system proposes a drone to land on a truck’s roof to re-supply it during operations; therefore, eliminating the need for the truck to return to the depot to gather more items to deliver. Thirdly, Murray and Chu (2015) discusses the benefits of utilizing drones to deliver relief items in disaster-affected areas. Drones avoid poor or non-existent infrastructure to deliver life-saving items to those in need. The paper focuses on an optimization model to deliver multiple light weight items in disaster areas and recharge stations to extend the operating range of the drones. Furthermore, Hong et al. (2017) introduces an optimization model that focuses on urban environments while also implementing recharging stations in the cities. Since drones are not permitted to fly over tall buildings or skyscrapers, route planning is essential for the operational efficiency of the drones. As the drones weave between buildings, their operational range reduces due to the extended distances needed to travel. However, many research articles do not discuss energy consumption within their articles. Research such as Dorling et al. (2016) fills in the gap by discussing vehicle routing problems with regards to energy consumption and hidden costs. They conclude that optimizing battery weight is an important variable to consider when evaluating optimal drone strategies. This strategy is utilized to supplement other cost and delivery time focused research. Lastly, Welch (2015) discusses the application of drone technology to be used for Amazon. A cost-benefit analysis is conducted on Amazon Prime Air to determine if drone delivery is more costeffective than traditional truck delivery. The paper notes current FAA regulations against commercial drone use and concludes that drone delivery is more costeffective than traditional truck. 3 Problem Statement Prescription drugs are life-saving medications that treat people with a variety of diseases, and, currently, prescriptions filling requires customers to come to the pharmacy to pick up the drugs. However, different regions have different infrastructure and variables that affect commute times for customers and can lead to customer dissatisfaction. While customer dissatisfaction may not be blamed on the pharmacies themselves, a pharmacy offering prescription drug delivery may attract new customers or increase customer loyalty. Belfast, Northern Ireland of the United Kingdom, was named UK’s worst city for traffic congestion in 2016 as citizens spend an extra 87% of their actual travel time stuck in traffic (Index, TomTom Traffic 2016). Therefore, Belfast provides a unique opportunity for pharmacies to service A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for. . . 189 their customers with prescription drug deliveries. On average, Northern Ireland pharmacies fill 315 prescriptions a day (United Kingdom 2019), and different delivery strategies are studied to explore possible methods for drug deliveries to enhance customer satisfaction. 4 Methodology 4.1 System Description A discrete-event simulation model is developed utilizing data gathered from pharmacies to accurately compare different methods for delivering prescription drugs. The model builds upon previous research and focuses on traditional Truck-Only, Truck-Tandem, and Drone-Only delivery methods. This model considers different scenarios in which customers do not travel to their local pharmacy, but the pharmacies provide delivery services for all their customers. However, healthcare is considered confidential and adheres to strict privacy laws. As such, patient data is closely guarded, and patient addresses are not provided by any pharmacy. Thus, certain assumptions are made. The prescription drop-off locations are randomized, but it is also assumed that the demand locations are around the pharmacy. Furthermore, prescription demand data is gathered elsewhere. The United Kingdom publishes pharmaceutical data on an open government license for public sector information. As such, pharmaceutical data of Belfast, Northern Ireland and United Kingdom is utilized to model the demand. Pharmacies operate on a 13-h workday, and the delivery vehicles are sent out 4 times a day as opposed to a continuous operation. The 13-h workday is similar to U.S. pharmacy operating hours, and the limited delivery operations ensure there is a batch of prescriptions that can be delivered at once. The pharmacies were found to fill, on average, 315 prescriptions a day. Therefore, the model demand follows a Poisson distribution with an arrival rate of 1/λ with λ = 315 prescriptions. Secondly, as the model compares different delivery methods, it is assumed that all trucks and drones are the same types with similar wear and tear. The drones are autonomous and do not require a human operator. Furthermore, the model does not consider battery or weight constraints on neither the truck nor the drone. However, energy, fuel, maintenance, and depreciation costs are considered for both the truck and drone. Lastly, the package release method from the drone is not considered in this model, and truck speed is reduced to reflect traffic on the roadways. Further, 1.5 min is given to each delivery to simulate the time it takes for the package to leave the vehicle and reach its destination. Figure 1 illustrates the process map the delivery vehicles must follow to deliver the prescriptions. 190 A. Jackson et al. Start Prescriptions are filled No Is it time for a batched to be released Yes Prescriptions are loaded onto vehicle Release drones to deliver Last-Mile Vehicle travels to demand point or cluster center point No Is location a demand point Yes Deliver Prescription Yes No Does vehicle hold more prescriptions Yes No Vehicle returns to pharmacy Are there prescriptions left to be delivered No Was the previous batch the last batch of the day Yes End Fig. 1 Process map of prescription delivery service A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for. . . 191 4.2 Sequence of Events 4.3 Data Collection and Analysis Pharmacy prescription demand is collected from the U.K. government (United Kingdom 2019). The data set provides monthly prescription orders from 2014 to 2018. A sample of demand during January from each year is taken and analyzed to provide average daily prescription orders and the growth trend shown in Fig. 2. Secondly, speed limits of Belfast are gathered, and congestion data are used to provide an average speed for a delivery truck in Belfast. The speed limit within the city is around 64.4 km/h, but citizens in Belfast spend an extra 87% of their time in traffic. Therefore, vehicles travel roughly half of the speed limit while in the city. Drones are not encumbered by speed limits, but by their speed capabilities. Data collected from previous research suggests that a Matternet drone can travel around 40 km/h (Scott and Scott 2017). Lastly, demand points are randomly generated within the simulation software Simio with the constraint being the range for the drones and the city for the trucks. Distances are calculated for the vehicles through the Simio Software. Table 2, below, shows the different parameters utilized in the model and can be replicated for further research. January Daily Orders 360 350 340 330 320 310 300 290 280 270 2014 2015 2016 Daily Orders Fig. 2 Daily orders in the month of January from 2014 to 2018 2017 2018 192 A. Jackson et al. Table 2 Model parameters Name Prescription arrivals Drone speed Truck speed Drone range Truck range Demand point locations Drone capacity Work day Batch distribution Delivery buffer Distribution Poisson Triangular Triangular NA NA NA NA NA Equally distributed NA Parameter λ = 315 prescriptions (0,40,72) km/h (0,32,72)km/h 10 km 386 km Randomly generated 1 package 13 h 4 batches a day 1.5 min 4.4 Model Parameters 5 Discrete Event Simulation Three scenarios are developed: Truck-Only, Truck-Tandem, and Drone-Only, to showcase possible delivery methods that also incorporate new technology. The necessary parameters are incorporated into the scenarios, and the scenario’s operational efficiency and effectiveness are compared with regards to batch delivery time and cost of implementation. The objective of the scenarios is to minimize the cost of a pharmaceutical delivery system while remaining capable of delivering the prescription batches with the given number of trucks, drones, and truck capacity being the decision variables. The baseline is established by taking the number of prescriptions to be filled on the workday and dividing it by the number of batch deliveries. A throughput constraint is placed to ensure all the prescriptions are delivered. The truck capacity variable is added to determine how loaded the trucks need to be to balance speed and economy. As shown in Table 3 Barnes and Langworthy (2004), Edwards (2011), Hawes (2016), Kim (2016), Rabta et al. (2018), Matternet: delivery drones that are delivering now (2015), United States Energy Information Administration (2019), United Postal Service (2017), Welch (2015), Figliozzi (2018), Eisenbach Consulting (2020), Wei and Figliozzi (2012), Australian Taxation Office (2019), costs are measured by capital, maintenance, labor, fuel, energy, and depreciation. Delivery Table 3 Cost comparison Investment cost Truck $50,000.00 Drone $5000.00 Maintenance ($ per km) $0.08 $0.03 Labor cost ($ per hour) $25.00 $0 Fuel cost ($ per Gallon) $2.69 $0 Energy cost ($ per kWh) $0 $0.03 Depreciation (% per year) 8.33% 50% A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for. . . 193 trucks depreciate at 8.33% a year, and drones lose roughly 50% of their value each year (Australian Taxation Office 2019), and so drones must be replaced frequently, increasing their costs. 5.1 Verification and Validation Multiple functions are in place to ensure that the model works properly. “Model trace” provides the live steps the model takes in its logic processing. The trace allows the user to follow along as the model runs its processes and can precisely pinpoint where errors occur. Secondly, a “pause” function is built within the model to pause the model whenever a minor error occurs. These errors only interrupt a small portion of the model, and the model can circumvent them to continue operation. Lastly, the model has a “stop” function that completely stops the model when a major error occurs. These errors affect the model as a whole and typically involve flaws in the logic processes. However, all errors can be pinpointed through the model trace, and a solution can be implemented accurately. Validation The t-statistic is used to determine a 95% confidence interval. A small random sample is taken from the pharmacy demand and is found to have a mean of 315 prescriptions a day with a standard deviation of 20.18. The 95% confidence interval is calculated to be 291,341 prescriptions in a day. A t-test (α = 0.05) was conducted to validate the demand arrival. The actual demand had a mean of 315 prescriptions a day, and the model had a mean of 309 prescriptions. This resulted in a p-value of 0.674, which is greater than 0.05, thus confirming that the demand is not statistically different. Given that pharmaceutical delivery and drone technology are relatively new, a t-test cannot be performed on delivery times as there is no real data to compare. 6 Alternative Scenarios 6.1 Scenario 1: Truck-Only As shown in Fig. 3, the demand points are randomly generated through the software within the limitations of the city. The trucks start from the pharmacy and take the shortest path to visit every demand point before returning to the pharmacy. Before that, the trucks wait for the prescriptions to be batched together. When a batch is ready, the prescriptions are loaded onto a truck, and the truck departs for 194 A. Jackson et al. Fig. 3 Illustration of demand drop off locations for scenario 1 delivery. The truck navigates the road to its appropriate demand points. Once the truck reaches a demand point, 1.5 min is allotted for the truck to make the delivery, and then the truck calculates the next demand point that needs service and sets out to make its delivery. The truck repeats the process until it has delivered all the prescriptions and then returns to the pharmacy. After returning to the pharmacy, the truck checks to see if more prescriptions in the batch need to be delivered; if not, the truck waits for the next batch to be ready. 6.2 Scenario 2: Truck-Tandem As shown in Fig. 4, the demand point clusters are randomly generated through the software within the limitations of the city. The trucks start at the pharmacy and wait for the prescriptions to be batched together to make their deliveries. When a batch is ready, the truck travels to a designated cluster center where it launches the drone. The drone makes the last-mile flight to the demand point to complete the delivery. Once the drone takes off, the drone flies above obstacles and in a direct line to a cluster’s demand points for quick delivery, where 1.5 min is allotted. The drone delivers the product at the demand location and returns to the truck. After the drone A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for. . . 195 Fig. 4 Illustration of demand point clusters for scenario 2 has returned to the truck, the truck proceeds to calculate the next cluster that it needs to serve. The truck determines the quickest route to the next cluster center point and travels by road to reach it. Upon reaching the center point, the drone delivery process is repeated until all demand points in the cluster are served. The truck continues this operation until it has delivered all its prescriptions and then returns to the pharmacy. After the truck has returned to the pharmacy, it determines if more prescriptions from the batch need to be delivered. If so, it will set out again to deliver the order, but if not, it will wait until the next batch is ready. 6.3 Scenario 3: Drone-Only As shown in Fig. 5, the demand point locations are randomly generated within the limitations of the drone range. The drones begin at the pharmacy and fly a direct 196 A. Jackson et al. Fig. 5 Illustration of the demand point for scenario 3 route to each demand point location to complete the delivery before returning to the pharmacy. Before leaving the pharmacy, the drones wait for the prescriptions to be batched together to make their deliveries. When a batch is ready, the prescriptions are loaded onto the drones, and the drones depart for delivery. Drones will be loaded only with 1 prescription to be delivered to a demand point. The drone then takes off and travels in a straight path to the demand point. Upon arrival, the drone will deliver the prescription and then return to the pharmacy to pick up another prescription. The process is repeated until all prescriptions are delivered in a batch. 7 Results The parameters of the models are implemented in Simio OptQuest with a warm-up period of 100 h. 300 simulations with 5 replications are used to achieve the results with a 95% confidence interval. OptQuest experimental setup in Simio considered 1–15 trucks with capacities ranging from 1–80 prescriptions for scenario 1, 1–15 A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for. . . 197 Table 4 Truck only results Truck only Number trucks 13 3 Number drones 0 0 Truck capacity 6 30 Investment cost $ 650,000 $ 150,000 Yearly cost $ 557,870 $ 364,607 Batch delivery time (h) 0.93 2.82 trucks with capacity ranging from 1–80 prescriptions and 1–5 drones for Scenario 2, and 1–30 drones for Scenario 3 to achieve its results. Note that the capacity of drones is considered as 1 prescription per delivery. 7.1 Scenario 1: Truck-Only As shown in Table 4, the delivery method consisting of 3 trucks with capacities of 30 prescriptions each is determined to be the cheapest feasible option. The batches are delivered in 2.82 h and require an investment of $150,000. However, the cost per year is $364,607, which includes labor, maintenance, fuel, and depreciation for the three trucks. 7.2 Scenario 2: Truck–Tandem As shown in Table 5, the delivery method with 1 truck, with a capacity of 79 prescriptions, equipped with 5 drones proves to be the cheapest feasible option with a batch delivery time of 2.17 h. This would require an investment of $75,000 and carries a yearly cost of $162,960, mostly due to the depreciation of the drones. 7.3 Scenario 3: Drone-Only As shown in Table 6, the delivery method with 4 drones proves to be the cheapest feasible option. The batch delivery time for 4 drones is 2.69 h and requires an Table 5 Truck-tandem results Truck-tandem Number trucks 7 1 Number drones 35 5 Truck capacity 12 79 Total investment cost $ 525,000 $ 75,000 Yearly cost $ 521,848 $ 162,960 Batch delivery time (h) 1.32 2.17 198 A. Jackson et al. Table 6 Drone only results Number of drones Drone only 30 4 Total investment cost Yearly cost Batch delivery time (h) $ 150,000 $ 113,558 0.39 $ 20,000 $ 49,602.5 2.69 investment of $20,000. However, yearly cost accounts for $49,602.5, mostly due to the depreciation of the drones. 7.4 Comparable Results As shown in Table 7, it is evident that the strategies can deliver prescriptions in a timely manner and offer cheap feasible options. However, the most effective option is also determined, given additional vehicles and capacity constraints. A DroneOnly strategy proves to be the most efficient. With enough drones, a pharmacy can deliver its first batch of prescriptions within 0.39 h while maintaining low yearly costs compared to other methods. Furthermore, the most cost-efficient method, Drone-Only delivery with 4 drones, only accrues an investment cost of $20,000 while delivering the batch in 2.69 h. It is important to note the significant drop in both capital and yearly costs that comes with the implementation of drones. However, Truck-Tandem is the most effective in delivering prescription drugs given the baseline. Using Simio OptQuest, the most effective strategy is also determined Table 7 Comparative results Number of Number of Truck Trucks Drones Capacity TruckOnly 13 0 6 Cost $ $ 650,000 557,870 $ $ 364,607 0 30 150,000 $ $ 7 35 12 525,000 521,848 $ $ 75,000 162,960 $ $ 113,558 1 5 79 DroneOnly Cost Yearly 3 TruckTandem Total Investment 0 30 0 150,000 $ $ 0 4 0 20,000 49,602 Batch Delivery Time (hrs) 0.93 2.82 1.32 2.17 0.39 2.69 A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for. . . 199 for each scenario to illustrate the cost increase and batch delivery time reduction with additional vehicles. 7.5 Sensitivity Analysis The initial analysis determined the least cost option while remaining feasible and the most effective options given additional vehicles. However, as demand grows with time, the impact on delivery time due to demand variation needs to be determined. Figure 6, below, shows the impact on batch delivery time for both the most effective and low-cost option for each scenario and includes a 3.25-h limitation to ensure that the daily orders can be completed. For the most effective option, variation below the current average has little to no impact. However, as the demand increases, batch delivery times for Truck-Tandem increase dramatically, and the current settings is no longer the most effective. The option now requires an adjustment in the number of trucks and truck capacity to become the most effective again. Furthermore, as the demand increases, the batch delivery times for all scenarios in the low-cost option increase steadily. However, both Truck-Only and Drone-Only become infeasible and will require an additional truck or drone. Lastly, as shown in Fig. 7, the impact on batch delivery times due to variation in the number of batches a pharmacy decides to operate with is determined. For the most effective option, an increase in the number of batches above the current setting has little impact. However, decreasing the number of batches, vastly increases the delivery times for both Truck-Only and Truck-Tandem. For the low-cost option, decreasing the number of batches vastly increases the delivery time for Truck-Only and Drone-Only, and will require additional trucks or drones to remain effective. However, increasing the number of batches, steadily decreases the batch delivery Low Cost Option Batch Order Time (hrs) Batch Order Time (hrs) Most Effective Option 3 2.5 2 1.5 1 0.5 0 215 265 315 365 415 4 3.5 3 2.5 2 1.5 1 0.5 0 215 Daily Demand Truck-Only Truck-Tandem 265 315 365 415 Daily Demand Drone-Only Truck-Only Truck-Tandem Drone-Only Fig. 6 Comparing the impact of daily demand variation on batch delivery times with the 3.25-h limitation for the 3 scenarios. The red line represents the upper limitation of batch delivery time to complete the daily delivery 200 A. Jackson et al. Low Cost Option 3 Batch Delivery Time (hrs) Batch Delivery Time (hrs) Most Effective Option 2.5 2 1.5 1 0.5 0 2 3 4 5 6 6 5.5 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 2 Truck-Only Truck-Tandem Drone-Only 3 4 5 6 Number of Batches Number of Batches Truck-Only Truck-Tandem Drone-Only Fig. 7 Comparing the Impact of the number of batches on batch delivery time with 3.25-h limitation between the 3 scenarios. The red lines illustrate the batch delivery time limitation for each number of batches times for all scenarios. Nevertheless, the low cost option for Truck-Only no longer becomes feasible as the trucks are no longer able to deliver the entire batch before the next delivery batch time. 7.6 Discussion In this section, the different scenarios are compared with respect to their cost, ease of implantation, performance, risk, and coverage. These comparisons are formed into a ranking system that a pharmacy manager can utilize for the decision-making process. As each scenario has potential benefits and problems, it is to the pharmacy manager to decide which is best for their pharmacy. Cost Each scenario requires costs to operate, and scenario 3 requires the least. Due to the relatively low costs of drones, scenario 3 is the cheapest option, followed by scenarios 2 and 1. Scenario 1 requires the most investment as trucks require a high investment and accrue higher costs as they are used. Scenario 2 is a combination of a truck and drone delivery system, and it serves as a middle ground between the two. Scenario 2 requires fewer capital expenditures as its assets are shifted from trucks and into drones. Furthermore, its costs are reduced as its labor and fuel costs are cut, and maintenance cost is reduced. Ease of Implementation Each scenario will require planning to implement, and scenario 1 contains the least number of obstacles to overcome. Trucks are vehicles that have been a part of society for decades, and a Truck-Only system will be the easiest to implement. Scenario 1 can be implemented almost immediately; however, scenario 3 will be the most difficult. As drone technology is relatively new, politics A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for. . . 201 and regulations offer substantial hurdles for companies to overcome. There are many concerns regarding drones that have led regulatory government bodies such as the Federal Aviation Administration to prevent the implementation of drones until further research could be completed. Scenario 2 again would serve as a middle ground between scenario 1 and 3 as it is a mixture of the two. However, scenario 2 will run into the same regulatory hurdles as Scenario 3; therefore, its ease of implementation is still behind Scenario 1. Performance Each scenario can deliver batch orders promptly; however, scenario 2 is the quickest with regards to its low-cost option that remains feasible. However, scenario 3 becomes the best performing if there are large amounts of drones. As drones are not limited by infrastructure or traffic, they can fly over many obstacles that would normally inhibit trucks. Therefore, scenario 3, with many drones, provides the quickest delivery times for a pharmacy. Due to the relatively low cost of operating drones, significantly more drones can be purchased, reducing the delivery times. Furthermore, scenario 2 is limited by space in the truck and cannot hold a large number of drones needed to increase its performance, but it will remain more stable than scenario 1 as the demand grows. Lastly, scenario 1 requires more trucks to be competitive with scenario 2. Risk Scenario 3 contains significant amounts of risk as drones are still developing, and the risk is exacerbated with additional drones. There is a safety risk as drones can cause mass damage if not operated appropriately, especially around airports. Furthermore, citizens are concerned about the risk drones pose to their privacy as drones are capable of circumventing traditional barriers. Moreover, it also faces a loss of property risk as armed citizens may destroy the drones they suspect of wrongdoing, or drones could malfunction, causing a crash. Political risk is abundant as regulatory agencies are unsure of allowing drones to operate. Lastly, operational risk is prevalent as technology and code are prone to malfunction, and packages may be delivered to the wrong location. However, many of these risks can be mitigated as technology and public education improves, and regulations become up-to-date. Scenario 1 bears the least amount of risk as trucks are a well-regulated technology and are confined to a predetermined path with its own regulations. However, scenario 1 bears some risk as driving conditions may be unpredictable and human error too is a factor. Scenario 2 bears the most risk as it contains both trucks and drones; therefore, combining the risks. Coverage Scenario 3 covers the least amount of territory as drones are limited in range. Due to their battery lives, drones are restricted to serve a small area around a pharmacy and require frequent recharges. Scenario 1 and 2 operate in a more extensive territory as the ranges are dependent on trucks. However, Scenario 2 covers the most range as it can launch a drone if a demand location is beyond the max range of a truck. 202 A. Jackson et al. Table 8 Rankings (3 being best to 1 being worst) Scenario 1: Truck-only 2: Truck-tandem 3: Drone-only Cost 1 2 3 Ease of implementation 3 2 1 Performance 1 2 3 Risk 3 1 2 Coverage 2 3 1 7.7 Managerial Implications The model scenarios provide insights into how a drone delivery service can be implemented at pharmacies. However, it is up to the pharmacy managers to determine which scenario is more suitable for them. Pharmacy managers are assumed to know what is best for their pharmacy, especially as they may have knowledge of unknown variables not incorporated into this model. However, rankings can be given for the manager to use to support them in decision making. As shown in Table 8, no scenario is dominant. Each scenario has its benefits and drawbacks, and it is up to the pharmacy manager to prioritize the rankings. Managers prioritizing a simple fix can utilize scenario 1 and avoid scenario 3. Scenario 1 requires the least number of obstacles to overcome, and scenario 3 will require the most effort to implement. Furthermore, if a manager requires the least amount of risk, scenario 1 can be utilized, and scenario 2 should be avoided. As trucks are the simplest to implement, they contain the least amount of risk. However, trucks bear some risk, and Scenario 2 adds that to the significant amount of risk drones carry. If a manager wishes to cover the most amount of territory, Scenario 2 can be utilized, and scenario 3 should be avoided. The combination of truck and drone allows Scenario 2 to cover the most territory, and the limitations of drone range confine Scenario 3 to a small area. Lastly, if a manager has a limited budget or wants the best performing system, Scenario 3 should be adopted, and Scenario 1 needs to be avoided. Drones perform the best within their range, and their low costs allow multiple drones to be purchased while remaining under budget. However, trucks require high capital expenditures and accrue high costs due to labor, fuel, and maintenance. 8 Conclusion and Future Work The application of drone technology to delivering pharmaceuticals can greatly benefit customers. In cities, like Belfast, with high levels of congestion, the introduction of drones can reduce the number of vehicles on the road by eliminating the need for people to travel to pharmacies to collect their prescription drugs. Furthermore, different methods for incorporating drones into delivery strategies offer different benefits. Three different scenarios: Truck-Only, Truck-Tandem, and Drone-Only, are researched and compared to determine the most optimal scenario A Simulation-Based Evaluation of Drone Integrated Delivery Strategies for. . . 203 with regards to delivery time and cost. The model suggests that a Drone-Only system is the most effective if there are a high number of drones, and the system still remains cheap and feasible. While each scenario can fulfill the demand, the different scenarios offer unique benefits and drawbacks that balance each other out. A pharmacy manager will be required to prioritize the rankings and determine which scenario is the best fit for them. This study has its limitation, and future work can improve the current model. Factors considering truck routing and drone features are not considered. Truck routing can be optimized with regards to the shortest path between demand points to calculate the true delivery times considering truck routing strategies. Furthermore, partnering with a pharmacy to gather real patient locations can make the model reflect real-world data more adequately. Next, factors such as package weight can be researched to determine how the weight will affect drone speed and battery life. Furthermore, different batteries can also be researched to determine the interval time between recharges. References Australian Taxation Office. (2019). TR 2019/5 - Income tax: effective life of depreciating assets (applicable from 1 July 2019). Barnes, G., & Langworthy, P. (2004). Per Mile Costs of Operating Automobiles and Trucks. Transportation Research Record, 1864(1), 71–77. Dayarian, I., Savelsbergh, M., & Clarke, J.-P. (2020). Same-day delivery with drone re-supply. Transportation Science, 54(1), 229–249. Deloitte. (2018). Deloitte study: Holiday Shoppers Hooked on fast and free shipping. Dorling, K., et al. (2016). Vehicle routing problems for drone delivery. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 47(1), 70–85. Edenhofer, A. (2018). DHL Parcelcopter, Deutsche Post DHL Group. Edwards, R. (2011). In struggle to cut expenses, hospitals eye the pharmacy. Hospitals & Health Networks, 85(11), 28. Eisenbach Consulting. Electricity rates by state (updated April 2020)”, © 2002–2017 Eisenbach Consulting, LLC. Epps, N., & Jordan, E. (2015). Coldwell Banker commercial affiliates survey reveals consumer expectations for same-day delivery, use of mobile wallet shaping retail experience, Coldwell Banker Commercial. Figliozzi, Mi. (2018). How sustainable are drone (UAV) deliveries? TREC Friday Seminar Series 149. Ha, Q. M., et al. (2018). On the min-cost traveling salesman problem with drone. Transportation Research Part C: Emerging Technologies, 86, 597–621. Hawes, C. (2016). UPS orders 200 workhorse group hybrid delivery trucks, Trucks.com. Hong, I., Kuby, M., & Murray, A. (2017). A deviation flow refueling location model for continuous space: A commercial drone delivery system for urban areas. Advances in Geocomputation, 2017, 125–132. Index, TomTom Traffic. (2016). Tomtom.com. Kim, E. (2016). The most staggering part about Amazon’s upcoming drone delivery service. Retrieved from http://www.businessinsider.de/cost-savings-from-amazon-dronedeliveries2016-6 204 A. Jackson et al. Matternet: delivery drones that are delivering now. (2015). Nanalyze.com. Retrieved December 3, 2015, from http://www.nanalyze.com/2015/12/matternet-delivery-drones-that-are-deliveringnow/.Published Murray, C. C., & Chu, A. G. (2015). The flying sidekick traveling salesman problem: Optimization of drone-assisted parcel delivery. Transportation Research Part C: Emerging Technologies, 54, 86–109. Rabta, B., Wankmüller, C., & Reiner, G. (2018). A drone fleet model for last-mile distribution in disaster relief operations. International Journal of Disaster Risk Reduction, 28, 107–112. ReportLinker. (2016). From warehouse to front door: Drones deliver the goods, Consumer Surveys, Latest Articles & Indicators, Transportation Industry. Scott, J., & Scott, C. (2017). Drone delivery models for healthcare. In Proceedings of the 50th Hawaii international conference on system sciences. United Kingdom. (2019). GP prescribing data, Government License. United Postal Service. (2017). UPS expands investment in vehicle collision mitigation technology, Press Room. United States Energy Information Administration. (2019). Gasoline and diesel fuel update. Wei, F., & Figliozzi, M. (2012). Conventional vs electric commercial vehicle fleets: A case study of economic and technological factors affecting the competitiveness of electric commercial vehicles in the USA. Procedia – Social and Behavior Sciences, 39, 702–711. Welch, A. (2015). A cost-benefit analysis of Amazon Prime Air. Zipline. (2020). How-it-works. Retrieved September 11, 2020, from https://flyzipline.com/how-itworks Pro-Active Strategies in Online Routing Stefan Bock In this paper, we summarize the basic ideas of recent approaches for the efficient controlling of urgent delivery processes in real-time. As in real-world applications the requests to be serviced the same day are usually unknown and will occur dynamically, such delivery processes possess a considerable degree of dynamism that has to be handled by a real-time approach. For this purpose, the transportation plan that is already in execution is continuously adapted by applying suitable optimization approaches according to all decisions that are not implemented. Due to the assumed urgency of the considered requests, the aim of these approaches is to minimize the total weighted request response times in order to minimize resulting customer inconveniences. Through the exploitation of past request data, stochastic knowledge about future request occurrences is derived that enables significant reductions of the request response times if the given data possesses a certain degree of diversity. Furthermore, empirical analyses reveal that the degree of structural diversity enables reliable forecasts concerning the positive impact of integrating stochastic knowledge derived from the given past request data. More recently, this approach has been extended for identifying recurring request patterns in real time in order to improve the quality of forecasts concerning future request arrivals. In addition to that, it has been shown that frequent en route diversions executed during iterative plan adaptations can be strongly limited without significantly worsening the attained request response times. Note that an unlimited exhaustive usage of en route diversions may cause an increase of accidents due to driver distractions. S. Bock () Business Computing and Operations Research, Schumpeter School of Business and Economics, University of Wuppertal, Wuppertal, Germany e-mail: sbock@winfor.de © Springer Nature Switzerland AG 2021 S. Srinivas et al. (eds.), Supply Chain Management in Manufacturing and Service Systems, International Series in Operations Research & Management Science 304, https://doi.org/10.1007/978-3-030-69265-0_8 205 206 S. Bock 1 Introduction This paper considers the well-known Vehicle Routing Problem (VRP). To the best of the author’s knowledge (see also Pillac et al. (2013)), the VRP was originally introduced as a generalization of the Traveling Salesman Problem by Dantzig and Ramser (1959). Usually, it is defined on a graph G = (V, E, C), with V = {v0 , v1 v1 , . . . , vn and the depot v0 ), , . . . , vn} (the set of all customer vertices E = vi , vj | vi , vj ∈ V × V ∧ i = j (the set of transportation edges), and C : E → R (the costs of using a transportation edge). For reasons of simplicity, the cost function C is frequently defined as a matrix Ci,j (with 0 ≤ i, j ≤ n and i = j) such that Ci,j gives the costs for traveling from vi to vj . In order to service all n customer vertices, a fleet of K homogeneous vehicles is available. Hence, the aim of the VRP is to find K cyclical tours that start and end at the central depot v0 while together visiting all customer vertices at minimum total transportation costs. The literature provides various extensions of this basic formulation (see Pillac et al. (2013) or Ferrucci (2013), pp. 32–38). For instance, the Capacitated-VRP (CVRP) extends the VRP by limiting the loading of each vehicle by imposing a maximum capacity while each customer has a predetermined demand. Moreover, the urgency or availability of some customers is mapped by integrating time windows at the customer vertices. As a consequence, a vehicle tour cannot service a customer before the opening of the respective window while servicing a customer after the end of the time window is either forbidden (hard time windows) or penalized (soft time windows). Hence, the respective variants of the VRP are denoted as the VRPTW or as the VRPSTW. This paper considers routing problems that deal with the delivery of urgent goods or services. Hence, material or immaterial goods (e.g., spare parts, repair services, specific technical support, or a combination of these examples) have to be delivered while an immediate delivery is of utmost importance to the customer. For instance, such deliveries become necessary after a customer complaint due to an unsuccessful first delivery or if this first delivery contains erroneous parts. In these exemplary cases, only an instant subsequent delivery of the needed goods or services in real time may limit the occurred customer inconvenience. Hence, in the literature these processes are frequently denoted as the distribution of perishable goods and the respectively considered real-world processes are named Real-Time Distribution of Perishable Goods (RDOPG) applications (see Ferrucci (2013), pp. 1–2). As a consequence, Ferrucci et al. (2013) propose the usage of a specifically defined objective function that pursues the finding of tour plans ζ minimizing the customer inconvenience z(ζ ). For each request i with a predefined weight wi , this customer inconvenience is operationalized as a function F(ti ) of the resulting request response time ti . This response time ti measures the time between request occurrence and request delivery. According to the results published by Davis and Maggard (1990) and by Kristensen et al. (1992), Ferrucci et al. (2013) model customer inconvenience as a linear as well as a quadratic function of the resulting request response times. The exemplary courses of both functions are sketched by Fig. 1. Pro-Active Strategies in Online Routing 207 Fig. 1 Illustration of the two customer inconvenience functions (linear2X and quadratic) modelled and analyzed by the approach of Ferrucci et al. (2013) (see Ferrucci (2013), p. 82) In order to additionally map a maximum response time Rmax that is communicated in advance to the respective customers, the inconvenience is significantly increased by Rpen after reaching this threshold. Thus, requests that are delivered after time Rmax are denoted as late requests. Furthermore, in order to ensure that those late requests are not delivered arbitrarily late in the linear case, the steepness of the increase of the linear objective function is doubled after reaching the threshold of the maximum response time Rmax . Note that this modification is not necessary for the quadratic case since this function weights each increase of the response time over-linearly such that late requests obtain a larger priority anyway. All in all, the objective function measuring the total inconvenience for delivering all non-serviced requests that constitute the set RτU is defined by: Minimize z (ζ ) = i∈RτU wi · F (ti ) + R pen · Θ ti − R max , with (x) = 1 for x > 0 and 0 otherwise see Ferrucci et al. (2013 p. 132). Another important issue, whose integration allows for covering various practical applications by the Vehicle Routing Problem, deals with the restricted or delayed availability of information. In contrast to the standardized deterministic VRP where all information is available, static, and reliable throughout the ongoing transportation process, dynamic variants of the VRP (abbreviated as DVRP) distinguish themselves by the fact that a considerable amount of data is not reliably given beforehand, but becomes available or changes during the execution of the transports. This complicates an efficient execution of the process as unexpected events like new requests or congestions may occur and have to be handled in real time and simultaneous to the execution of the ongoing transports. As illustrated in Fig. 2, the service of dynamically incoming requests has to be integrated into the existing vehicle tours that are already in execution. For this purpose, the existing and 208 S. Bock Fig. 2 Illustration of the Dynamic Vehicle Routing Problem with newly incoming requests Fig. 3 Possible information flow within a real-time control approach (see Bock (2010), p. 734) already executed tour plan has to be adapted. Note that the significant progress of communication and information technologies makes it possible to control a fleet of vehicles in real time with a centralized approach that is continuously informed about the current situation in the network in order to handle incoming events by necessary plan adaptions. These adaptations are in turn communicated to the respective vehicles. The possible information flow within a real-time control concept is depicted by Fig. 3. The task of the central database component is to map the current situation of the controlled transportation process. For this purpose, recent events affecting the state of the vehicle fleet or of the transportation network are transferred by respective services and are directly integrated into the mapped situation. As a result, a static optimization problem is derived that defines the current state of the ongoing process by respective parameters (e.g., current vehicle positions, vehicle states, or known requests), whereas all changeable decisions constitute the variables of this model. In order to handle incoming events and to ensure an efficient process execution, the Pro-Active Strategies in Online Routing 209 real-time control has to continuously adapt the plan that is already in execution. Note that the concurrency of plan execution and adaptation requires a sophisticated update handling as – due to the complexity of the considered optimization problem – plan adaptations cannot be done in zero time, but require substantial computational time. One main characterizing aspect of each DVRP-instance that significantly drives the complexity of efficiently controlling the considered routing process is the ratio of unexpectedly occurring events and the time pressure under which these events have to be handled. Depending on the specific setting of the respectively considered DVRP, the literature provides different so-called degrees of dynamism (dod) measuring this aspect. According to Lund et al. (1996), the dod is originally defined by δ = nd /ntot . In this computation, nd is the number of unexpected dynamically occurring requests and ntot defines the total number of all requests (static and dynamic ones) to be satisfied by the tour plan. In order to additionally integrate the time pressure under which the dynamically incoming requests have to be handled, Larsen (2000) extended the dod δ to the effective dod that is defined by ria nd a δe = i=1 T ·ntot . In this calculation, ri ≤ T gives the arrival time of request i, while T provides an upper bound of all request arrival times. If, in addition, time windows exist, Larsen (2000) recommends applying the effective dod with time windows δ etw instead. This measure sums up the difference between T and the remaining response time before the closing of the respective time window ril T − r l −r a 1 d i i over all dynamically incoming requests, i.e., it holds δ et w = ntot · ni=1 . T Hence, this definition measures the time pressure for handling each event by the maximum remaining time before reaching the end of the respective time window. 2 A Reactive Real-Time Approach This section briefly sketches a reactive real-time approach that is also used (in a slightly modified variant) for the pro-active concept introduced later. It applies a basic concept that has been evaluated in terms of being able to efficiently control various dynamic, unreliable real-world processes in production and logistics (see Gendreau et al. (1999), Ichoua et al. (2000), Bock et al. (2006), Bock (2010), Ferrucci and Bock (2014)). Before starting the delivery process, a first transportation plan (denoted as the first relevant plan P0r ) for servicing the initially known requests is generated by applying a suitable optimization procedure. Subsequently, at time 0, this first relevant plan P0r is set in execution. In what follows, we denote Pτr as the relevant plan executed at time τ . As the Vehicle Routing Problem is an N P-hard problem, the finding of efficient tour plans is a complex task and cannot be done in zero or almost zero time. Hence, simultaneous to the execution of the ongoing transportation process, the real-time control approach continuously tries to improve all not implemented decisions of a currently stored theoretical plan Pτt . During this improvement process, the best performing theoretical plan Pτbt is kept for a possible 210 S. Bock exchange with the current relevant plan. In other words, the entire execution time of the ongoing process is used for improving all still theoretical plan decisions. Due to the concurrency of plan execution and possible plan adaptations, the real-time concept ensures a correct plan update by iteratively solving static snapshots of the ongoing process within predetermined time limits. These time limits are denoted as anticipation horizons and separate the continuous time lapse into discrete intervals of customizable length ta . Specifically, with the beginning of a new anticipation horizon at time τ , all requests that have occurred and buffered during the preceding horizon are integrated into the three stored plans, namely into the relevant plan, the theoretical plan, and into the best theoretical plan. Subsequently, the best performing plan of these three plans is taken as the new relevant plan. In order to ensure that the theoretical transportation plan that is modified by the applied optimization procedure during the considered anticipation horizon starting at time τ is applicable after the elapse of this horizon at time τ + ta , all decisions that will be taken by the relevant plan during this horizon are pre-simulated. Hence, after this pre-simulation, the theoretical plan only contains decisions that will be taken after time τ + ta . Therefore, the optimization procedure is applied to a static problem instance that maps a situation in the delivery network that will occur in the future, namely at time τ + ta . As a consequence, with the beginning of a new anticipation horizon, all decisions to be taken during this new horizon become fixed and are assumed to be taken as defined by the relevant plan. For further details of the applied plan adaptation methods, we refer to Bock (2010) and Ferrucci (2013). Figures 4 and 5 illustrate the described update handling of a Fig. 4 Continuously adapting the current relevant plan by iteratively solving static snapshots of the ongoing process at the adaptation level (cf. Bock 2010) Pro-Active Strategies in Online Routing 211 Fig. 5 Continuously adapting the current relevant plan by iteratively solving static snapshots of the ongoing process at the adaptation level (cf. Bock 2010) sophisticated reactive real-time control approach proposed in Bock (2010). Figure 4 depicts the situation in which new requests have arrived during the last anticipation horizon. These pending requests are iteratively integrated into the three stored plans by finding least cost insertion positions. As these modifications can be conducted in almost zero time, the modified plan can be instantly set in execution. However, as these instant modifications and the respective cost consequences are frequently of limited quality, this plan is subsequently adapted by applying the Tabu Search approach. Hence, as indicated by Fig. 5, these subsequent optimization processes often attain significant improvements of the current relevant plan. Note that, as the focus of this paper is solely on the design of the presented real-time approaches, all details of the applied Tabu Search procedure are neglected. For these important algorithmic details, we refer the reader to Ferrucci et al. (2013) and in particular to Ferrucci (2013, pp. 185–203). As Bock (2010) shows, a continuous optimization of the current relevant plan may result in significant improvements also in comparison to an event-oriented variant that solely conducts a plan adaptation by applying the insertion heuristic followed by a Tabu Search algorithm for the subsequent anticipation horizon after the occurrence of a dynamic event (e.g., the occurrence of a new request, a street congestion or street blockage, or technical vehicle problems). In contrast to this, Ferrucci (2013) reports that both approaches attained comparable results for the tested dynamic instances of the Vehicle Routing Problem. These differences can 212 S. Bock be mainly attributed to the fact that the problem instances of the extended Pickup and Delivery Problem considered in Bock (2010), which integrates transshipments and external service providers, are considerably more complex. Therefore, more substantial adaptations are frequently necessary to attain efficient process executions and to allow for an adequate handling of the various incoming events. This makes a continuous application of the optimization process reasonable. In contrast to this, most of the considered dynamic VRP-instances were efficiently adapted in a single adaptation horizon by the applied Tabu Search algorithm. However, Ferrucci (2013) reports that the continuous optimization outperforms the event-oriented one for larger sized instances. In order to reduce the time effort necessary for conducting the various computational simulations, all real-time simulations conducted by Ferrucci et al. (2013) apply solely the event-oriented variant. Note that the reactive real-time approach applies a so-called idle vehicle waiting strategy. This strategy requires that each idle vehicle, i.e., a vehicle that does not possess an assigned non-serviced request, has to wait at its current location of the network until a new request is assigned to be serviced. As a consequence, if the majority of requests occur regularly in specific local areas, this positioning of idle vehicles is quite reasonable as future requests can be expected close by their current locations. If, however, a vehicle possesses further requests in its tour (determined by the current relevant tour plan) this vehicle starts traveling to the next request location directly after completing the service of the preceding one. 3 Exploiting Past Request Data and Building the Stochastic Knowledge Empirical tests reveal that the aforementioned reactive real-time approach is able to ensure an efficient execution of the routing processes even if a high degree of dynamism has to be handled. This can be ascribed to the fact that by continuously optimizing all remaining decisions of the current relevant plan in execution, this approach is able to attain the necessary adaptability. As a consequence, increased response times, caused by necessary instant plan adaptations conducted due to incoming events, can be frequently mitigated due to a later finding of improved plans (see Bock (2010), Ferrucci (2013), pp. 220–224). Specifically, Ferrucci (2013) reports that, due to the fast improvements attained by the applied Tabu Search procedure, its optimization process can be restricted to 10 s after each dynamic event as, after this time limit, the applied meta heuristic is able to attain almost optimal tour plans. However, although these results are promising, closer analyses also reveal that sometimes vehicles have to conduct considerable detours to service dynamically incoming requests that are located far away while the tours of other vehicles have visited these areas earlier. Hence, if it would be possible to forecast some future Pro-Active Strategies in Online Routing 213 requests and integrate this information into the tour plan, response times may be further reduced. For this purpose, Ferrucci et al. (2013) propose to exploit available past request data in a preceding offline step that derives forecasts of future requests. By doing so, the reactive real-time approach is turned into a proactive one. However, in contrast to many other pro-active approaches, this method does not assume or apply any knowledge about existing distributions or patterns defining the dynamism of the routing processes without proving their existence. Rather, stochastic knowledge is derived solely from past data while all the required attributes needed for deriving the respective forecasts are checked. 3.1 Derivation of Stochastic Knowledge by Applying a Specifically Designed Cluster Analysis In order to derive stochastic knowledge about the occurrence of future requests, past data is separated into segments. As illustrated in Fig. 6, each segment s represents all incoming requests that occur in a quadratic spatial area with side length DCse during a specific time interval of length DCte . The incoming requests of each segment are modeled by a time-space Poisson process ps with a request arrival rate λ(ps ) that is calculated by the average number of requests arriving in the spatial-time assignment of segment s during the past nf days. Ferrucci et al. (2013) report that the applied setting DCse = 2.5 km and DCte = 1 minute leads only to low rate values λ(ps ). These low rates are frequently not significant enough to justify the rerouting of a vehicle at a specific time to the defined spatial area. Hence, after generating the segments, neighboring segments are iteratively combined into so-called clusters in order to fulfill given restrictions. Specifically, clusters are generated by the combination of segments that are located close to each other such that the following criteria are met: • The cluster does not exceed a predetermined maximum spatial (DCmse ) and temporal extension (DCmte ), as otherwise, its forecast information is not significant (the occurrence area is too large). • The sum of the rate values of the assigned segments is at least DCminλ , i.e., a minimum total request occurrence probability is guaranteed in this cluster. • Further quality criteria are fulfilled (see, for details, Sect. 3.2). Fig. 6 Definition and combination of neighboring segments 214 S. Bock Clearly, one may assume that the finding of such clusters can be done by applying one of the well-known standardized clustering approaches such as, for instance, one k-means clustering variant (see Arthur and Vassilvitskii (2007) or Jain et al. (1999)). However, these approaches require the predetermination of the number of clusters to be built. Moreover, the fulfillment of the itemized quality measures cannot be controlled in these standardized approaches. Hence, Ferrucci et al. (2013) develop a new specifically designed clustering method that solves a defined Mixed Integer Problem (MIP) by applying the CPLEX standard solver. In order to map the entire set of possible clustering alternatives, all clusters that are theoretically generatable and fulfill the defined quality criteria are built and integrated into the defined MIP. Note that this includes overlapping and therefore mutually excluding clusters. 3.2 Ensuring Two Further Cluster Quality Criteria After being generated, each (still theoretical) cluster ctemp is checked as to whether it fulfills the following two quality criteria: (i) a maximum average travel time inside the cluster that is necessary for servicing the respective requests and (ii) that the request arrivals fulfill, at least to a customizable extent, basic requirements of a nonhomogeneous time-space Poisson process. Note that the fulfillment of the second criterion is assumed and used throughout the approach. In what follows, some details concerning both criteria are depicted (cf. Ferrucci (2013), pp. 175–177): To criterion (i): As the actual location of a dynamically incoming request may differ from the derived cluster location, travel times inside a cluster are an issue of assessing the quality of a generated cluster. Namely, as the intention of deriving a cluster is to pro-actively guide a vehicle to the corresponding request-likely area at the specified time in order to reduce request response time, the travel time inside a cluster has to be kept strongly limited. For this purpose, the connectivity of the barycenter (derived for the requests occurring in considered cluster ctemp during the last nf days) inside the real-world street network is checked first. This is done by applying the Dijkstra algorithm (cf. Dijkstra (1959)) to find, within a predetermined travel time radius DCradiusTT , the nearest location on a road of a class possessing a maximum speed limit (i.e., a road categorized in the street network allowing for the fastest travel). If such a location cannot be found within the prescribed time distance, a location in this area is chosen instead that is located on a road with maximum speed limit. In both cases, this node is denoted as nctemp and represents the location the corresponding dummy customer is assigned to if the respective cluster is chosen for being integrated into the stochastic knowledge. After determining the location nctemp (i.e., the avgT T location that vehicles are pro-actively rerouted to) the average travel time ct emp is computed to all past request locations that occurred in ctemp within the last avgT T nf days. If ct emp exceeds the predefined threshold DCmaxAvgTT , the temporary cluster ctemp is discarded. Pro-Active Strategies in Online Routing 215 To criterion (ii): Each generated potential cluster (that may become a dummy request) is checked as to whether it represents basic attributes of a nonhomogeneous time-space Poisson process, at least, to a certain extent. Note that these attributes are used for the updating of the remaining occurrence probabilities of future requests in the generated clusters that are needed by the applied Tabu Search approach to weight the dummy requests throughout the ongoing computations. For this purpose, N(t) is denoted as a stochastic counting process that measures the incoming requests in the considered cluster ctemp until time t. According to Ross (2019), the following three criteria have to be checked for N(t). First, it holds that N(0) = 0. As this criterion is obviously fulfilled, it does not have to be checked for cluster ctemp . Second, for time length t, N(t) is Poisson distributed with mean t · λ(ctemp ) as λ(ctemp ) denotes the rate parameter of the Poisson process. Third, the numbers of incoming requests in disjoint time periods are independent. In order to check the second criterion for a possible cluster ctemp, Ferrucci et al. (2013) apply the Chi-Square Goodness-of-Fit test with a customizable type I error α according to the guidelines described by Kvam and Vidakovic (2007, p. 156). Specifically, by this test, it is analyzed whether substantial deviations exist between the actual number of occurred requests in the respective cluster ctemp during the past nf days and those that would be expected by applying a Poisson distribution with the derived parameter λ(ctemp ). For this purpose, the sample X1 , . . . , Xnf is considered as the number of observed request arrivals in cluster ctemp during the past nf days. For the details of this test, the reader is referred to Ferrucci (2013, pp. 175–177). The authors applied the Chi-Square Goodness-of-Fit test for different values of α. Although in the literature very small values of α = 0.05 are frequently proposed (cf., e.g., Israel (2008, p. xxii) or Sapsford and Jupp (2006, p. 217), Ferrucci (2013) reports that, in the considered application, these quality thresholds were much too low to guarantee the presence of Poisson process characteristics with a sufficiently high reliability. Hence, best results were achieved by applying the threshold α = 0.4. This value leads to an average Poisson quality in the clusters of more than 0.6 for the generated data sets as well as for the real-world data. Ferrucci (2013) states that although further increasing α improved this value, setting α to 0.4 guaranteed an efficient compromise between the conflicting goals of ensuring an acceptable minimum Poisson quality in the clusters and reducing erroneous rejections of clusters that possess a sufficient Poisson quality. The third criterion of Poisson processes that requires that the requests occur independently for disjoint time intervals (cf. Ross 2019) can also be checked by using an appropriate ChiSquare Goodness-of-Fit test. However, due to the small temporal extension of a cluster, Ferrucci et al. (2013) refrain from performing this test. This is because the required accuracy in request arrival times is frequently not fulfilled by practical request data sets and therefore would lead to an unfavorable rejection of clusters (for further details, see Ferrucci (2013, pp. 175–177). Only if both checks are successfully completed, is the considered cluster ctemp kept for being integrated into the MIP formulation for finding a maximum number 216 S. Bock of non-overlapping clusters. In other words, this cluster is a candidate for generating a later dummy request that has be integrated into some vehicle tour plan by the reactive real-time approach. All those candidates constitute the cluster set C. The potential dummy request that arises from the considered cluster ctemp is characterized by the following parameters: W eight of the dummy request : wi = P (X ≥ 1) = 1 − P Sλ(ctemp ) (0). avgT T Service time of the dummy request : si = R st + ct emp · λ ct emp . As P Sλ(ctemp ) (0) provides the probability that no request will occur in ctemp , the defined weight-parameter wi gives the probability that at least one request will occur in ctemp. As this parameter discounts the contribution (the customer inconvenience as a rated response time) of the request, this weight parameter gives the priority of the dummy request within the reactive real-time control. Moreover, the derived service time for traveling and servicing the expected number of requests in ctemp is used for estimating the time that a vehicle is kept at the respective location of the derived dummy request nctemp of cluster ctemp. 3.3 Selection of the Clusters As stated above, the preceding steps have led to the constitution of a set of all potential (not necessarily non-overlapping) clusters C that are considered for being chosen to define additional dummy requests to be additionally integrated into the tour plans by the reactive real-time approach. In order to compute a suitable set of non-overlapping clusters out this set C, the following MIP is defined. In this MIP, the binary variable xi for i ∈ C is used that is one (xi = 1) if and only if the cluster i ∈ C is selected and zero otherwise. Furthermore, the parameter cist art T L gives the temporal threshold at which cluster ci starts and M defines a big number dominating all other values. M· xi − cist art T L · xi Maximize i∈C s.t. (1) ∀s ∈ SC : i∈C ci,s · xi ≤ 1 and (2) xi ≤ C max. i∈C The above MIP pursues the maximization of the number of generated clusters. However, if this number is identical, solutions with earlier starting clusters are preferred. Restriction (1) ensures that chosen clusters are non-overlapping. This is ensured by using the parameter ci, s that is one if and only if segment s ∈ SC belongs to cluster i. Hence, if there is some overlapping in the cluster selection, there exists a segment that is assigned to more than one chosen cluster. This is excluded by restriction (1). Moreover, restriction (2) defines an upper bound of the number of Pro-Active Strategies in Online Routing 217 minλ chosen clusters. This bound is initially set to C max = λ /DC (p ) s s∈SC with SC defining the set of all segments belonging to the clusters of set C. Thus, a maximum number of clusters can be obtained by the total number of expected requests over all segments divided by the minimum number of expected requests required for building a cluster. The bound defined in restriction (2) is used for technical reasons as it enables the applied CPLEX solver to derive tighter bounds. These tighter bounds enable a reduction of the solution space that has to be explored by the conducted examination process. By solving the MIP through the application of the CPLEX solver, the set of generated clusters is determined. Each of these clusters defines a dummy request with the aforementioned parameter setting that is added to the set of real requests. Therefore, it has to be integrated into the tour plans computed and controlled by the reactive real-time approach. 4 Transforming the Reactive Real-Time Approach into a Pro-active One The applied reactive real-time approach is able to attain a high adaptability that handles incoming requests in an efficient way even when a high degree of dynamism is given. However, response times of urgent requests can be further reduced if these requests can be reliably forecasted and this stochastic knowledge is integrated into the tour scheduling process. For this purpose, the generated clusters allow for the additional definition of so-called dummy requests possessing a specific parameter setting that distinguishes them from ordinary requests. While ordinary requests have the weight one (occurrence is deterministic), each dummy request is discounted with its occurrence probability riw ≤ 1 (defined in Sect. 3.2). Each ordinary request has a uniform fixed service time of 60 s, whereas the service time of a dummy request is set to the total time necessary on average for servicing all requests that are expected in the respective time-spatial area (see Sect. 3.2). Each dummy request i being defined by cluster ci is integrated into the set of requests that have to be delivered by the reactive real-time control approach. However, while the ordinary requests require a delivery (or a pick-up) of the respective goods (or persons), the service of a dummy request by the reactive approach has a modified intention. It reroutes some vehicle to the defined spatial area at the specified time cist art and keeps it there until time ciend . Both dummy request dependent parameters result from the segments belonging to ci . Note that the integration of dummy requests requires a special handling by the reactive real-time control approach during the execution of the transportation process. As long as this derived interval is not reached by the current time τ + , all aforementioned parameters of the dummy request i remain unchanged. Specifically, if τ + ≤ cist art holds, we have wi (τ + ) = wi and si (τ + ) = si . If, however, the end of the current anticipation horizon τ + exceeds cist art , i.e., if it holds that τ + > cist art , the parameter values of the corresponding dummy requests 218 S. Bock change dynamically with the lapse of time. For this purpose, Ferrucci et al. (2013) apply the following update handling for the dummy requests: The beginning of the time window (ei ): ei τ + = cist art τ + + ciend − cist art τ + · ∞ 1 n=1 P Sλ(ci ,τ + ) (n)· n+1 1 − P Sλ(ci ,τ + ) (0) The idea behind this updating scheme is to compute at time τ + the expected arrival time of the first future request if altogether n = 1, 2, . . . , ∞ additional requests will occur in the remaining segments (i.e., assigned segments that start at time τ + or later) of cluster ci . If n > 0 requests occur, there are n + 1 intervals before and after these arrivals that are assumed to be equally distributed over the 1 entire remaining interval. Thus, ∞ n=1 P Sλ(ci ,τ + ) (n)· n+1 gives the proportion of the remaining interval ciend − cist art τ + that is expected to elapse until the first request arises. As this calculation assumes that at least one additional request will occur, it has to be divided by the probability that no request occurs, namely, it has to be divided by 1 − P Sλ(ci ,τ + ) (0). By multiplying the sum of the constant service time Rst and the average travel avgT T time ci with the number of requests still expected at time τ + in the cluster ci , + st avgT T the updated service time si is obtained. si τ = R + ci · λ ci , τ + As each dummy request is weighted (i.e., is discounted) with the probability that at least one additional request will occur in the remaining segments of cluster ci and this probability continuously changes when τ + > cist art holds, this weight wi is updated by the formula: wi τ + = P (X ≥ 1) = 1 − P Sλ(ci ,τ + ) (0) A dummy customer i is finally removed from the set of requests to be serviced when it becomes unlikely that a further request will arrive in the remaining segments of the considered cluster, i.e., if λ(ci , τ + ) < DCλrem holds. Clearly, this time point tcrem can be computed in advance. i The integration of dummy requests requires the application of an extended waiting strategy that is briefly described in what follows: Whenever a dummy request i becomes the next request that has to be serviced in a vehicle tour, it may be possible that, due to the defined time window, this service is in the far future. In such a situation, a direct visit of dummy request i is not reasonable as the respective vehicle would arrive much too early in a remotely located area and therefore would not be available for other, currently more urgent requests, in central regions. Therefore, in such a situation, the travel to the defined location of a dummy request is delayed until the point in time cist art minus the necessary travel time to the location nci of the dummy request i is reached. In other words, the vehicle waits at its current location until the last point in time when an instant travel to nci allows Pro-Active Strategies in Online Routing 219 for a timely arrival at nci . Clearly, during this waiting time, the respective vehicle is available for servicing other requests. 5 Computational Evaluation In order to evaluate the proposed approach, Ferrucci et al. (2013) conduct a series of various computational experiments. The proposed approach was tested for two different request data classes, namely, SREAL and SGEN . While the daily instances of SREAL stem from a real world subsequent delivery process of a German newspaper publishing company, all experiments of the request data class SGEN were additionally generated according to the main request occurrence characteristics of SREAL. Note that the considered subsequent delivery process is characterized by a high dynamism as on average there are 150 request arrivals per day, while these dynamic arrivals mainly occur during the 4 h between 7 and 11 am. Hence, the simulation of each day conducted by Ferrucci et al. (2013) solely maps the process of this most challenging 4 h period. At first, this section introduces the generation process of the data class SGEN . Subsequently, the measured results of the pro-active approach for the real world data class SREAL and for the various test sets of the generated data class SGEN are presented and analyzed. 5.1 Generating the Instances of the Data Class SGEN All experiments simulate the transportation process in real time on the realworld road network of the city of Dortmund (a medium-sized German city). For this purpose, a discrete-event based simulator was developed and applied. The algorithms were implemented in Delphi and the experiments were conducted on Personal Computers equipped with Pentium D 2.8 GHz CPU and 2.5 GB Ram. The considered region covers an area of 22.5 km × 20 km = 450 km2 . In order to test different scenarios with individual capacity demands, fleet sizes of 8, 10, and 12 were simulated. At the beginning of each experiment, an initial tour plan was generated by applying the Tabu Search procedure for 120 seconds. This plan assigns all known real requests and dummy requests to vehicle tours. In what follows, we describe the main aspects of the instance generation process that is applied by Ferrucci et al. (2013) to constitute the set SGEN . Note that this generation is comparable to the method applied by Ichoua et al. (2006). The used parameter values in all generated instances of SGEN are empirically derived in preliminary tests and given in Table 1. In order to simulate various settings with individual distributions of dynamically incoming requests, the considered service area is divided into a set of P disjunctive uniformly-sized quadratic subregions denoted as a1 , . . . , aP . Moreover, the time lapse is separated into Q time slices of equal uniform length. For each defined 220 S. Bock Table 1 Values of the main parameters applied during the computational experiments Parameter nf DCse × DCse DCte DCmse DCtse DCradiusTT DCmaxAvgTT DCminλ DCλrem RD TD Q P nd /ntot Srd Rpen Rmax F(Rmax ) Brief description Number of past days used Spatial extension of each segment Temporal extension of each segment Maximum spatial cluster extension in both directions Maximum temporal cluster extension Maximum travel time radius within a cluster Maximum average travel time radius within a cluster Minimum occurrence probability of a request controlled by a predetermined minimum λ-value Threshold for removing a request RegionDiversity TimeDiversity Number of time slices Number of subregions Proportion of dynamically incoming requests Size of each request data set Penalty value for late requests Maximum response time Largest inconvenience value for a non-late request Value/values 60 days 2.5 km × 2.5 km 1 min 2 segments 15 min 300 s 650 s 1.0, 1.2, 1.5,1.8,2.0 (i.e., probabilities are 0.6321, 0.6988, 0.7769, 0.8347, 0.8647) 0.25, 0.5 (i.e., probabilities are 0.2212, 0.3945) 0, 0.25, 0.5, 0.75, 1.0 0, 0.25, 0.5, 0.75, 1.0 5 18 0.9 30 instances 100 3600 s 1.0 subregion, a specific time-space Poisson process is applied to generate the arrival times and locations of incoming requests. Specifically, the parameter λ(ai , tj ) gives for subregion ai in time slice tj the respective request arrival rate. Hence, in order to generate incoming requests for each time interval tj , the corresponding sum λ tj = P i=1 λ ai , tj provides the time-dependent arrival time of the next request. The subregion where the request occurs (i.e., is assigned to) is randomly drawn while the probability of choosing subregion ai amounts to p(ai, tj ) = λ(ai , tj )/λ(tj ). Finally, the position within subregion ai is randomly drawn. In order to obtain individual streams of random numbers, extended versions of the random number generator proposed by Park and Miller (1988) and Park et al. (1993) are applied (see, for further details, Ferrucci (2013, pp. 211–217). The main reason for separating the spatial area into P subregions a1 , . . . , aP stems from the cognition that the performance of the applied pro-active elements of the approach are substantially dependent on the occurring variances within the request arrival patterns over the simulated days. In other words, the positive impact of using the stochastic knowledge exploited from the request arrivals of the past days Pro-Active Strategies in Online Routing 221 Fig. 7 Linear averaging exemplarily applied to the individual request arrival probabilities within the subregions a1 , . . . , aP . (see Ferrucci (2013), p. 213) depends on the extent by which the spatial assignments of dynamically incoming requests change throughout the day. In order to be able to analyze this important issue, the aforementioned setting allows for suitable customizations of the request data within SGEN . For this purpose, Ferrucci et al. (2013) introduce the terms RegionDiversity (RD) and TimeDiversity (TD) that measure the structural diversity of the dynamically incoming requests. Specifically, TimeDiversity specifies the variance in each defined subregion over the time lapse that is separated into Q time slices t1 , . . . , tQ . Analogously, the RegionDiversity gives the variance between different subregions for each time slice. In order to systematically generate different instances with various diversity levels, the generation of SGEN starts with a setting that possesses maximum diversity values. This is attained (and therefore defined) by giving each subregion ai in each time slice tj an individual request arrival rate λ(ai , tj ). This initial scenario is characterized by the parameter setting TD = 1.0 and RD = 1.0. Further settings are generated by reducing these diversities in predefined steps through the application of linear averaging. As illustrated in Fig. 7, in each step of this process, the variety of the assumed request occurrence probabilities either for a given subarea along the time lapse or for a given time slice between different subareas is reduced by restricting the interval between maximum or minimum deviations from the average value. Specifically, in case of reducing RD, a time slice is kept and the variety of the defined request occurrence probabilities of the subregions is reduced. In case of reducing TD, the same is done for a kept subregion along the time lapse. As shown in Fig. 7, this averaging is done such that altogether five different states of RegionDiversity and TimeDiversity arise. These are denoted as RD/TD = 1.0, RD/TD = 0.75, RD/TD = 0.5, RD/TD = 0.25, and RD/TD = 0 By combining these five states, we obtain altogether 25 settings for evaluating the impact of the existing diversity within the request occurrences on the performance of the proactive approach. However, note that the RD- and TD-measures do not define the 222 S. Bock structural diversity in a general way, but allow for a relative comparison of generated instances. While instances of set SGEN generated with the setting RD/TD = 0 do not possess any diversity (i.e., the dynamism does not change spatially nor with the time lapse) a considerable dynamism can be expected from instances of set SGEN generated with the setting RD/TD = 1.0. As depicted in Table 1, during the experiments, the number of subregions P was set to 18 while the number of time slices Q was set to 5. Further details of the instance generation process executed with one of the 24 settings are given by Ferrucci (2013, pp. 211–217). 5.2 Measured Results for the Instances of the Data Class SREAL The proposed reactive and the pro-active real-time approach were applied to 30 instances of the real-world data class SREAL. For comparison reasons, a Greedy heuristic was also applied that handles dynamically incoming requests by finding a least cost insertion position in the currently executed transportation plan. Note that this heuristic is also applied by the sophisticated reactive real-time approach whenever new requests occurred during the preceding anticipation horizon. However, to find further plan improvements, the reactive real-time approach subsequently applies the Tabu Search heuristic for 10 seconds. Table 2 gives the average improvement rates attained by the reactive real-time approach in direct comparison with the greedy heuristic. Comparable to the results reported by the studies of Gendreau et al. (1999), Ichoua et al. (2000), and Bock (2010), the reactive approach substantially outperforms the greedy heuristic by between 9% (fleet of 12 vehicles) and 32.45% (fleet of 8 vehicles) for the linear objective function and by between 19% (12 vehicles) and almost 52% (8 vehicles) for the quadratic objective function. These improvements underline that an efficient control of such kind of dynamic processes requires the application of a sophisticated real-time approach that is able to continuously adapt Table 2 Improvements attained for data class SGEN by the reactive real-time control approach in direct comparison with the greedy heuristic according to the average customer inconvenience and the number of late requests Objective function Linear2X 8 vehicles 10 vehicles 12 vehicles Objective function quadratic 8 vehicles 10 vehicles 12 vehicles Improvement vs. Greedy heuristic 32.45% / 11 vs. 175 late 14.33% / 0 vs. 9 late 9.21% / 0 vs. 0 late Improvement vs. Greedy heuristic 51.96% / 14 vs. 158 late 30.65% / 0 vs. 5 late 19.28% / 0 vs. 1 late Pro-Active Strategies in Online Routing 223 Table 3 The average number of clusters that are attained by applying CPLEX for solving the MIP to optimality (see Sect. 3.1) in dependence of the tested DCminλ -values (see Ferrucci (2013, p. 232) Number of clusters generated in dependence of the quality threshold DCminλ DCminλ = 1.0 DCminλ = 1.2 DCminλ = 1.5 DCminλ = 1.8 DCminλ = 2.0 67.7 44.6 24.4 13.1 9.4 the ongoing transportation plan according to the incoming dynamic events. This is particularly true if, due to a smaller fleet size, the number of available vehicles is strongly limited. Obviously, such adaptability is not reachable by applying a simple least cost insertion greedy heuristic. In order to additionally apply and evaluate the pro-active approach, stochastic knowledge has to be exploited from SREAL in an offline step as described in Sect. 3.1. For this purpose, aside from the 30 days used for the performance assessment, a further 60 days are selected. These exploitations were repeated with different threshold values of DCminλ determining the minimum request occurrence probability in a cluster. For these evaluated DCminλ -values, Table 3 gives the average number of clusters (and therefore dummy requests) generated by applying CPLEX. It is worth mentioning that the finding of optimal solutions of the MIP are reasonable as these solutions provide a substantially increased number of generated clusters. Ferrucci (2013) reports that, in direct comparison with an alternatively applied heuristic, optimal solutions of CPLEX provide between 7.7% (for DCminλ = 2.0) and 13.4% (for DCminλ = 1.0) more clusters. Note that the number of generated non-overlapping clusters constitutes the stochastic knowledge of a predetermined quality exploited from the given data set. Therefore, the improvements attainable by the application of the pro-active approach solely depend on them. By analyzing the results attained by the pro-active real-time approach in comparison to the reactive one (see Table 4), no clear dominance can be identified. The pro-active approach uses the stochastic knowledge efficiently only in cases of larger fleet sizes (10 or 12 vehicles) and by applying the quadratic customer inconvenience objective function. In these cases, and in contrast to the linear customer inconvenience function, the quadratic customer inconvenience function prevents larger delays of urgent requests in favor of dummy requests that require larger detours. Moreover, due to the larger fleet sizes, falsely forecasted requests have a milder effect as the available vehicle capacity is not that scarce. However, the positive effects of pro-actively guiding vehicles to uncovered areas where forecasted requests actually occur, were observed only rarely. This can be ascribed to the fact that further analyses of the instances of the data set SREAL reveal that these instances possess a limited diversity. For instance, Ferrucci (2013) shows that the request arrival rates in the 72 subregions are quite stationary in so far that the subregions obtaining the majority of incoming requests remain quite stable during the day. Hence, as the reactive approach keeps idle vehicles at their last customer location, scenarios are quite unlikely where these positions are far away from future request occurrences. Due to this limited 224 S. Bock Table 4 The average improvement rates attained by the pro-active approach in comparison with the reactive one Results of the pro-active real-time approach in comparison with the reactive one forSREAL Vehicles Predetermined minimum quality of stochastic knowledge (Parameter DCminλ ) 1.0 1.2 1.5 1.8 2.0 Linear customer inconvenience objective function 8 −11.5% (28) ↑ −7.66% (21) ↑ −3.12% (18) ↑ −0.27% (18) ↑ 1.6% (10) ↑ 10 −4.86% (22) −1.42% (20) −0.68% (17) 1% (16) 0.01% (16) 12 −0.67% (15) −0.14% (16) 1.7% (13) 0.64% (11) 0.88% (11) Quadratic customer inconvenience objective function 8 −19.91% (27) ↑ −7.52% (21) ↑ 0.21% (14) ↑ 1.17% (15) ↑ 1.13% (13) ↑ 10 −3.99% (20) 0.7% (12) 2.76% (9) 2.79% (10) 0.71% (14) 12 6.59% (9) 7.93% (5) 6.78% (9) 7.31% (8) 6.08% (8) Values in bold print are the results for the best choices of DCminλ . The value in brackets gives the number of instances with worse results. Entries with “↑” indicate that the number of late requests increased by applying the pro-active approach (see Ferrucci (2013), p. 236) diversity in SREAL , the attained improvements of applying the pro-active approach are strongly limited. In order to analyze this aspect in more detail, the data class SGEN was additionally considered. 5.3 Measured Results for the Instances of the Data Class SGEN A first indicator of the substantial impact of a varying diversity in the considered data set on the resulting performance of the pro-active approach is provided by Table 5. This table gives the number of generated clusters for 5 of the 25 generated settings of RegionDiversity and TimeDiversity combined with five values of DCminλ . Note that these clusters represent the stochastic knowledge that can be exploited form the past data in order to extend the reactive real-time approach. Specifically, as indicated by Table 5, for each tested DCminλ value, it can be observed that the number of exploitable clusters significantly increases for data sets that possess a larger diversity. Table 5 Number of generated dummy requests in dependence of the level of generated diversity and the required occurrence probability DCminλ (see Ferrucci et al. (2013)) Diversity level RD TD 0 0 0.25 0.25 0.5 0.5 0.75 0.75 1 1 DCminλ 1.0 1.2 2 0 17 3 36 20 54 38 80 65 1.5 0 0 7 24 36 1.8 0 0 0 15 35 2.0 0 0 0 11 29 Pro-Active Strategies in Online Routing 225 Fig. 8 The attained improvement rates of the pro-active approach for the linear customer inconvenience objective function and a fleet size of 10 vehicles Fig. 9 The attained improvement rates of the pro-active approach for the quadratic customer inconvenience objective function and a fleet size of 10 vehicles This can be ascribed to the fact that the increased time and spatial variety in the data allows for the finding of clusters that fulfill the required quality issues. Hence, if, for instance, the maximum values DCminλ = 1.8 or DCminλ = 2.0 are applied, no cluster is found for the settings of lower diversity RD = TD = 0, RD = TD = 0.25, and RD = TD = 0.5. In contrast to this, clusters of this higher quality are only found for the diversity settings RD = TD = 0.75 and RD = TD = 1. In accordance with these results are the measured improvement rates attained by the pro-active real-time approach and highlighted by Figs. 8 and 9 for the two simulated objective functions, respectively. These two figures indicate that the pro-active approach outperforms its reactive counterpart for all 30 instances of SGEN if the full diversity (generated by 226 S. Bock the setting RD = TD = 1) is given. Specifically, an average improvement of 22.95% for the linear customer inconvenience function and of 39.33% for the quadratic customer inconvenience function can be observed. These improvements are also quite stable and significant for the diversity settings where both parameters possess a value of at least 0.75. Note that while Figs. 8 and 9 provide only the results attained by using a fleet size of 10 vehicles, comparable results are attained for 8 or 12 vehicles (see Ferrucci et al. (2013), p. 138). However, it is worth mentioning that a reduction of the available fleet size reduces the attained improvement rates of the pro-active approach, whereas an increase of the available vehicles leads to even higher improvements. Specifically, for the setting RD = TD = 1 with the highest diversity, the attained improvement rates with 8 vehicles are 19.51% in the linear case and 29.86% for the quadratic customer inconvenience function. But, if 12 vehicles are available, these improvement rates increase to 26.72% in the linear case and for the quadratic case even to 47.24%. The observed impact of the size of the fleet can be explained by the fact that in case of scarce resources the negative consequences caused by a falsely rerouted vehicle are more substantial. Consequently, it is reasonable to require larger DCminλ -values for these settings. Conversely, if a larger fleet size is available, it is promising to reduce the minimum cluster quality (i.e., by defining smaller DCminλ -values) to some extent. Hence, owing to the measured results, it can be concluded that there are two main criteria for a successful integration of the derived stochastic knowledge into the reactive real-time approach (cf. Ferrucci et al. (2013), p. 137). The first criterion is a sufficient diversity in the request data. As indicated by the measured results, this diversity enables the finding of a significant number of clusters of a minimum required quality. Hence, the presence of a sufficiently high level of diversity allows for the generation of reliable stochastic knowledge about future request occurrences. Moreover, due to the variety that is significant for highly diversified request data, this knowledge has a substantial value for the real-time distribution process. This results from the fact that the given variety of the request arrivals reduces the efficiency of the vehicle positioning strategy applied by the reactive approach. When idle vehicles wait at the location of the last service, a large variety within the occurrences of future requests requires that these vehicles have to be frequently guided to areas that are located at a substantial distance. This underlines the value of possessing reliable knowledge about these locations where a future service is needed at the beginning of the process. The second criterion that substantially influences the efficiency of replacing the reactive approach by the pro-active one is the available fleet size. As mentioned above, larger fleet sizes are best handled by applying smaller DCminλ -values as the negative impacts of incorrectly forecasted dummy requests diminish with a more relaxed capacity situation. Conversely, for smaller fleet sizes, it is reasonable to require a larger minimum cluster quality. Therefore, depending on the size of the available fleet size, a suitable value for the parameter DCminλ has to be derived. One further significant finding of the study of Ferrucci et al. (2013) is the definition and evaluation of the degree of structural diversity (the dosd) as a general offline measure. It allows for reliable estimations of the practicability of Pro-Active Strategies in Online Routing 227 Fig. 10 Illustration of the computation of the degree of structural diversity (dosd) integrating stochastic knowledge into the reactive real-time control in order to reduce the resulting request response times. The basic idea of this measure is to derive a normalized value that reveals the movement of the spatial barycenter of the incoming requests over time. For this purpose, the daily time lapse is divided into T time periods of m minutes each. For each time period t ∈ {1, . . . , T}, the y spatial barycenter of nt incoming requests is computed as a tuple btx , bt ∈ Gx,y of geographical positions according to the x- and y-axis with the domain of possible positions G = {(g1 , g2 ) 0 ≤ g1 ≤ x ∧ 0 ≤ g2 ≤ y }. Hence, the dosd measure sums up the distances between the resulting barycenters over all time periods (see Fig. 10). Each distance is normalized by dividing it by the trivial upper bound value maxDist = x 2 + y 2 . Consequently, we obtain the following mathematical definition of the dosd that is illustrated in Fig. 10: dosd = T −1 t =1 nt +1 · ! 2 y y 2 btx+1 − btx + bt +1 − bt maxDist· T −1 t =1 nt +1 Note that this definition omits the movement position of the barycenter to the y of the first time period, i.e., the movement to b1x , b1 . This is done as the initial position would be the central depot where all vehicle tours start from. Therefore, this first time period is used to obtain instance-dependent vehicle positions. But, as the number of requests occurring during the first period n1 is kept small in the considered experiments, the omission of the first movement does not have a significant influence on the computed dosd-values (for details, see Ferrucci (2013), pp. 249–254). Ferrucci et al. (2013) evaluate the practicability of the proposed dosd by analyzing the attained improvements rates for all conducted experiments (possessing a comparable number of serviced requests) depending on the dosd- 228 S. Bock value calculated for the respectively given request data set. In order to keep the numbers of serviced requests per vehicle comparable, these analyses were generated for different sizes of the used vehicle fleets. In order to assess best or almost best results for all experiments, the highest improvement rates over all tested parameter settings (i.e., the most suitable value for DCminλ ) were taken as the respective result. In order to derive the correlation between the attained improvement rates and the dosd of the request data, Ferrucci et al. (2013) apply six linear regression models for each of the applied inconvenience functions. By applying the Pearson product-moment correlation coefficient r according to the description given by Rodgers and Nicewander (1988), convincing values between 0.9161 and 0.9647 were attained depending on the fleet sizes or assumed inconvenience functions (linear or quadratic). This proves a strong correlation between the dosd and the reduction of the resulting response time attainable by using stochastic knowledge exploited from the available request data. Ferrucci et al. (2013) therefore conclude that the diversity in a request data set with a dosd ≥ 0.6 enables the proposed pro-active approach to attain considerable reductions of the request response time. Clearly, these conclusions also depend on the applied exploitation methods (applied for finding the clusters) and on the definition of the dosd. But, as the reactive approach keeps idle vehicles at the location of their last service, a successful use of pro-active vehicle relocation strategies basically depends to some extent on a minimum existing diversity in the data. Therefore, it can be stated that the derived dosd constitutes a suitable offline measure that improves the applicability of the proposed real-time approach. Namely, it allows reliable conclusions for suitably customizing the pro-active methods and whether significant reductions of the attainable response times of the reactive real-time approach are possible by integrating stochastic knowledge concerning future request occurrences. For further details, we refer to the more detailed evaluations and descriptions provided by Ferrucci (2013). 6 Efficiently Controlling en Route Diversions Many real-time approaches that adapt an existing tour plan simultaneous to its execution use en route diversions. Such an en route diversion defines a situation where a vehicle is diverted from its current travel to a request destination by a plan modification conducted and communicated in real time. This en route measure is possible due to advances in communication technologies integrated in the exemplary information flow of a real-time control depicted by Fig. 3. Specifically, these systems allow for locating vehicles in real time as well as for an instant communication of new orders to the drivers (see Giaglis et al. (2004), Larsen et al. (2008), or Ferrucci and Bock (2015)). By doing so, the adaptability of a realtime control and therefore the efficiency of the tour execution can be considerably improved. For instance, if the delivery or pick-up location of a newly incoming request is in the direct vicinity of the position of a vehicle currently traveling to Pro-Active Strategies in Online Routing 229 another request destination, an instant diversion of this vehicle may significantly reduce the response time of the respective request if other vehicles are positioned farther away. Thus, due to these positive effects, to the best knowledge of the author of this paper, aside from the approach of Ferrucci and Bock (2015), no real-time control approach considers limiting the usage of en route diversions. However, while the positive consequences of en route diversions are quite obvious and lead to considerable efficiency improvements, an exhaustive application may also have negative consequences. First of all, vehicles have to be equipped with the technical equipment required while drivers have to be taught how to operate the devices in an efficient way. But, despite the possible effort invested in the training of drivers, numerous en route diversions to be conducted under the tight time restrictions of a real-time control are still quite demanding. Hence, the resulting stress level of the driver may cause substantial visual and/or cognitive distractions with longer and more frequent off-road glances (see Owens et al. (2010) or Kaber et al. (2012)). This, in turn, could be a significant cause of accidents (see Ferrucci and Bock (2015) or Young et al. (2013)). Ferrucci and Bock (2015) (see p. 81) distinguish between three different kinds of en route diversion that may occur during the application of the proposed pro-active real-time approach. 1. Diversion from a real customer (abbr. Dr ): While traveling to the location of a real request either another real request or a dummy request is assigned to the considered vehicle, or a waiting command is received. The latter forces the vehicle to wait at the current position for further instructions. Note that this covers a switching between different dummy requests. 2. Diversion from a dummy customer (abbr. Dd ): In this case, the travel to a dummy customer location is stopped due to the assignment of a real request. 3. Diversion from an outdated dummy customer: This special case occurs if a vehicle is traveling to the location of a dummy customer request and in the meantime, this request becomes outdated as the remaining occurrence probability of a request falls below the predetermined threshold. Note that, due to the absence of dummy requests, the second two cases cannot occur when applying a reactive real-time approach. Moreover, the first two cases are actively brought about by plan modifications while the last case occurs by the removal of dummy requests. Since these removals are known in advance, drivers can be informed to avoid the aforementioned negative consequences. As the third kind of en route diversion occurs only rarely, it is not considered further by Ferrucci and Bock (2015). In order to rate the remaining two kinds of en route diversions, two different driver profiles are considered. The first driver profile is denoted as a “job enrichment” orientation (abbreviated as JE) and assumes that a driver prefers to service real requests. As a consequence, such a driver appreciates changes from a dummy request to a real request. Hence, only the en route diversions of the first type negatively affect such a driver. In contrast to this, drivers of the profile “change avoidance” are negatively affected by all unforeseen changes (abbreviated as CA). 230 S. Bock Thus, both kinds of en route diversion are interpreted as negative and should be kept limited. In order to be able to control the number of conducted en route diversions in the real-time approach, a penalty cost function D(ζ ) is integrated into the pursued inconvenience objective z(ζ ). By charging a customizable penalty rate p for each conducted en route diversion, a considered tour plan ζ that causes n(ζ ) en route diversions has to account for an additional penalty D(ζ ) = n(ζ ) · p. Note that n(ζ ) is computed dependent upon the assumed driver profile (JE or CA). As a consequence of this extension, the reduced costs attained by the best of the two tour plans Pτt and Pτbt that are generated during a considered anticipation horizon (starting at time τ ) has to compensate for the additional costs being charged for a potentially increased number of planned en route diversions. Otherwise, the real-time approach does not change the relevant plan Pτr that is currently in execution. In other words, depending on the customizable cost rate p, additional en route diversions are only integrated into the tour plan if the resulting cost reductions are substantial enough that they exceed these penalties. The attained cost reductions are reductions of the resulting response time and can be computed depending upon the given objective function (linear or quadratic) and the defined cost rate p (see Table 6). This allows for a customizable control of using en route diversions through determining a threshold of necessary response time reductions. For practical applications, the cost rate p may be derived from a required minimum reduction of the resulting response time that should compensate for the each additionally caused en route diversion. In order to evaluate suitable values for p and to analyze the consequences of resulting en route diversions and customer inconvenience values, Ferrucci and Bock (2015) conduct various computational experiments. For a detailed depiction of the measured results, we refer the reader to the paper of Ferrucci and Bock (2015). However, the main results of this study can be made clear by considering Tables 7 and 8 that show the effect of the different penalty cost rates p for the two simulated objective functions (linear and quadratic, respectively) and the driver profile “change avoidance”. Table 6 Response time thresholds and penalty cost rates for en route diversions (see Ferrucci and Bock (2015), p. 82) Response time threshold (s) 0 (unrestricted case) 36 72 180 360 720 1800 3600 Complete prohibition Value for p dependent upon the pursued objective function Linear Quadratic 0 0 0.01 0.0001 0.02 0.0004 0.05 0.0025 0.1 0.01 0.2 0.04 0.5 0.25 1.0 1.0 1,000,000 1,000,000 Pro-Active Strategies in Online Routing 231 Table 7 Percentage changes of the total linear customer inconvenience attained in comparison to the unrestricted case by applying various values for the penalty cost rate p > 0 and using a fleet size of 10 vehicles Profile CA Obj. val. #Dr #Dd Distance Values p=0 0.2698 54.2 27.6 609,074 Percentage reductions/increases with the penalty cost rate p= 0.01 0.02 0.05 0.1 0.2 0.5 1.0 1,000,000 −1.4 −0.6 0.4 2.2 8.1 18.2 25.6 27.2 −23.4 −33.5 −57.9 −74.1 −87.6 −97.8 −99.8 −100 −8.8 −15.1 −27.3 −37.4 −61.4 −90.1 −98.1 −100 0.2 0.4 1.9 4.3 6.1 8.4 9.6 9.7 The simulated driver profile is CA Table 8 Percentage changes of the total quadratic customer inconvenience attained in comparison to the unrestricted case by applying various values for the penalty cost rate p > 0 and using a fleet size of 10 vehicles Profile Values CA 0 Obj. Val. 0.094 #Dr 52.3 #Dd 30.1 Distance 620,405 Percentage reductions/increases with the penalty cost rate p= 0.0001 0.0004 0.0025 0.01 0.04 0.25 1.0 1,000,000 0.1 0 −0.1 2.0 3.7 29.7 56 61.8 −1 −0.5 −12.4 −36.4 −69.8 −96.4 −99.8 −100 −0.7 −5.8 −2.3 −12.5 −26.0 −76.7 −98.5 −100 0.2 −0.4 0.2 1.8 3.3 7.6 9.8 9.8 The simulated driver profile is CA By considering these results for both mapped inconvenience functions, it becomes obvious that small penalty cost rates do not significantly increase the resulting customer inconvenience, but substantially reduce occurring en route diversions. All in all, due to the attained results, Ferrucci and Bock (2015) state the following managerial implications depending on three different practical settings: • Setting 1—The minimization of customer inconvenience is the only objective to be considered: Hence, in this case, one may deactivate the penalty cost extension. However, throughout the measured results, it can be observed that the smallest penalty values may result in further decreases of the resulting customer inconvenience. Hence, applying the en route diversion cost rate with those values is reasonable. • Setting 2—The minimization of both objectives (customer inconvenience and the number of en route diversions) are relevant: In this case, the application of moderate penalty values seems to be useful. For instance, by applying a penalty value p = 0.05 in the linear case and p = 0.04 in the quadratic case, the number of occurring diversions is reduced by more than 50%. Simultaneously, the customer inconvenience is only slightly increased. • Setting 3—The minimization of the number of occurring en route diversions is the only objective to be considered: In this case, it is reasonable either to apply the prohibitive cost rate p = 1,000,000 or to consider using a still large penalty value of 1.0. In direct comparison with the prohibitive value, the latter substantially 232 S. Bock reduces the resulting customer inconvenience, but allows only a very moderate number of diversions. Due to the derived cognitions of the computational evaluation, it can be stated that the extension of the real-time approach proposed by Ferrucci and Bock (2015) allows for a suitable control of en route diversions for the first time. Hence, the aforementioned negative consequences of this measure can be suitably covered. 7 Identifying Multiple Profiles in the Past Request Data A second considerable extension of the pro-active real-time approach is proposed in the paper by Ferrucci and Bock (2016). It allows for the identification of multiple request patterns in the past data in order to improve the quality of the derived stochastic knowledge and its efficient application. In contrast to this, the previously considered approaches (Ferrucci et al. (2013) and Ferrucci and Bock (2015)) assume that the request occurrences on each day follow a more or less identical pattern. As a consequence, no distinction of groups of past days is done and the approaches derive a single profile from the past request data to anticipate future request arrivals. In contrast to this, by grouping similar past days, the approach of Ferrucci and Bock (2016) allows for the identification of more than one profile. Specifically, it is assumed that the request occurrences of each day follow specific attributes that stem from an unknown set of patterns. As illustrated in Fig. 11, these patterns are not observable, but recur over time in a random fashion while influencing the request arrivals on the respective day. Hence, as patterns cannot be directly observed and identified, Ferrucci and Bock (2016) try to derive characterizing attributes of the structure of days by grouping similar past days. This is done by applying a suitable clustering algorithm. For this purpose, similarity between days and/or groups of days is defined by a specifically defined distance measure. By separating the time lapse of each day into time slices of predetermined length (for details, see Ferrucci and Bock (2016)), this measure calculates the total sum over all these time slices according to the following three quadratically weighted criteria: 1. Euclidean distance between the barycenters of occurred requests, i.e., for each time slice, the barycenter of requests occurring during this time period is computed for each day or group of days. Subsequently, the quadratic sum of the Euclidean distances between these barycenters of the compared (groups of) days is generated. 2. Difference of the number of occurred requests, i.e., for each time slice, the quadratic difference of the number of requests occurring during this time period is computed. 3. Difference of the average request distances to the barycenter, i.e., for each time slice, the distribution of requests occurring during the respective time period within each day (or group of days) is compared. For this purpose, the Euclidean Pro-Active Strategies in Online Routing 233 Fig. 11 Illustration of the basic assumptions of the multi-pattern approach distance of each request occurring in the respective time period to the respective barycenter of the day (or group of days) is computed. In order to compute the sought profiles as groups of similar past days, the k-means++ approach proposed by Arthur and Vassilvitskii (2007) is applied. By using a sophisticated seeding, this approach provides a competitive heuristic solution to this NP-hard clustering problem within a short time limit. Moreover, by incorporating some randomness, it produces various profiles in repeated runs. Hence, Ferrucci and Bock (2016) apply this clustering algorithm 1000 times. Subsequently, the best assignment is selected that possesses the minimum total distance of days to their assigned profiles. The main idea of the approach of Ferrucci and Bock (2016) is to improve the quality of the derived and applied stochastic knowledge by filtering out the past days that belong to a profile whose request occurrence structure is similar to the one observed for the ongoing day. In other words, dummy requests should be generated solely from those past days that are assumed to possess an identical or comparable pattern as that observed for the current day. Hence, after generating the profiles in an offline step, the real-time approach of Ferrucci and Bock (2016) analyzes the structure of incoming requests in order to identify similar known profiles, i.e., the respectively assigned days, in real time. Only these days should be used to derive suitable dummy requests. However, due to the significant computational complexity of the MIP that has to be solved for deriving these dummy requests (see Sec. 3), this computation is done beforehand in an effortful offline step. Therefore, after generating the respective profiles as sets of past days, the dummy requests are computed for each subset of existing profiles. Specifically, for each subset of 234 S. Bock profiles, to derive the respective dummy requests, the stochastic knowledge building procedure is applied that is described in Sect. 3. Hence, for k existing profiles, this procedure is applied 2k − 1 times by the offline procedure. Clearly, this exponential complexity limits the number of profiles which can be generated. One may think that a significant drawback of applying a k-means clustering algorithm for identifying existing profiles in the data set consists of the fact that this approach requires a predetermination of the number of profiles k to be built. Indeed, this is a significant prerequisite of the clustering algorithm. However, after starting with a two-group clustering by setting k :− 2, this can be handled by iteratively increasing the parameter k by one (setting k :− k + 1) until the new clustering with k + 1 groups no longer significantly outperforms the former one possessing only k groups. In order to identify in real-time the subset of profiles that are relevant in time slice t (note that these time slices are identical with the time slices used in the profile definition) for the ongoing day, Ferrucci and Bock (2016) propose two different rules: • Standardized Rule (SR): This rule keeps a profile p ∈ P as relevant if its total distance to the current day over the elapsed times slices (i.e., the time slices 1, . . . , t) does not exceed by more than 20% the maximum distance of a day within profile p to the profile itself over the time slices 1, . . . , t. I.e., SR keeps a profile for the current day at time slice t if it holds that Δ (d c , p) < 1.2· Δmax t,p = 1.2· maxΔt (i, p). In this computation, the abbreviation Dp gives the set of days i∈Dp assigned to profile p ∈ P and t (i, j) defines the total sum of distance measure values for the periods 1, . . . , t for two sets of days i and j. Note that a factor of 1.2 is assumed to cover some additional acceptance tolerance. • Extended Rule (ER): While the standardized rule rates all profiles independently of each other, the extended rule applies a rating that is relative to the profile that best matches the current day at the considered time. For this purpose, at first, for the periods 1, . . . , t, a best matching profile p∗ is identified, i.e., a profile fulfilling Δt (d c , p∗ ) = minΔt (d c , p). Then, accept the subp∈P # " with pt = set of profiles p ∈ P | Δt (d c , p) < (1.5 − 0.05· t) · pt · Δmax t,p max c ∗ Δt (d , p ) /Δt,p∗ . The extended rule becomes restrictive if profile p∗ matches a current day well as, in this case, pt becomes small. This fast orientation towards a well matching profile is the main intention of the rule. By subtracting 0.05 · t max pt · Δmax t,p from the threshold 1. 5 pt · Δt,p , this orientation is strengthened when the day unfolds. Ferrucci and Bock (2016) evaluate the new approach by various computational experiments. As one intention of these experiments is to assess the ability of the proposed approach to identify underlying multiple patterns, the data class SGEN is no longer suitable. As a consequence, the days of a newly generated data class are built according to predetermined patterns, while these patterns (i.e., specifications that function as templates for the generation of days with prescribed subregion- Pro-Active Strategies in Online Routing 235 dependent request arrival probabilities) are built analogously to the days of the data class SGEN introduced in Sect. 5.1. Following the findings of the study of Ferrucci et al. (2013) concerning specific structures of request arrivals supporting a successful application of the pro-active approach, all patterns are generated such that they possess a high diversity. In other words, it is ensured that the regions that possess the most incoming requests change from time slice to time slice. Hence, the arrival probabilities of the 18 subregions differ significantly from time slice to time slice, while the expected request arrival rate in each time slice is identical in all patterns. Subsequently, by applying each generated pattern, 60 days are respectively built. In what follows, the abbreviations 2p, 3p, and 4p are used and denote a data set of 120 days based on two patterns (2p), a data set of 180 days based on three patterns (3p), and a data set of 240 days that is based on four patterns (4p). In what follows, the original approach (proposed in Ferrucci et al. (2013)) that uses only a single profile is identified with an extension “SP”, while the extended version that applies 2, 3, 4, or more profiles obtains the extension “MP”. In order to decide about the decision rule used for filtering out the relevant profiles in real time, Table 9 gives the improvement rates attained by replacing the pro-activeSP-approach by the pro-active-MP approach for different settings of the number of generated patterns, assumed profiles, and applied objective functions (linear or quadratic customer inconvenience). These improvement rates indicate a substantial outperformance of SR by ER that is further increased for settings being generated by using more patterns. This clear dominance of ER can be attributed to the fact that ER is able to identify the relevant profiles much faster than SR. Specifically, by comparing each profile with the best matching profile, potentially existing outliers in some profiles do not harm the identification process. This can be particularly observed while applying SR to settings where the number of profiles coincides with the number of patterns (i.e., if k = 2 is applied to 2p, if k = 3 is applied to 3p, or if k = 4 is applied to 4p). Table 9 Averagely attained improvement rates by applying the extended profile acceptance rule (ER) (instead of SR) within the pro-active approach for 2, 3, or 4 profiles, best DCminλ -values, and a fleet size of 10 vehicles Number of patterns 2p 3p 4p 2p 3p 4p Number of generated profiles k=2 k=3 k=4 k=5 Linear objective function – objective values 4.8% 4.5% 3.9% 3.1% 8.8% 12.5% 4.1% 4.0% 2.0% 7.7% 12.7% 8.9% Quadratic objective function – objective values 12.6% 9.9% 8.0% 13.2 18.4% 25.0% 12.2% 12.0% 4.7% 16.0% 23.5% 18.1% k=6 3.5% 2.7% 9.2% 9.3% 10.5% 19.0% Values in bold print are the largest improvement rates attained by ER for two, three, and four patterns, respectively (see Ferrucci and Bock (2016), p. 367) 236 S. Bock Table 10 Average improvement rates in comparison to the standardized aproach with a single profile attained by applying k = 2, 3, 4, 5, and 6 profiles for the linear inconvenience objective function (see Ferrucci and Bock (2016), p. 366) Linear Number of vehicles 8 10 12 Number of generated profiles Applied patterns 2p 3p 4p 2p 3p 4p 2p 3p 4p k=2 6.7% 5.2% 0.7% 7.0% 10.0% 0.9% 8.6% 10.7% 0.8% k=3 5.0% 7.6% 4.2% 6.0% 13.9% 6.8% 8.8% 17.5% 9.4% k=4 4.4% 8.0% 7.6% 6.0% 14.4% 12.4% 7.2% 16.9% 17.6% k=5 6.2% 9.3% 6.9% 7.5% 13.9% 11.9% 9.2% 18.0% 16.3% k=6 4.4% 7.8% 9.3% 6.4% 13.5% 12.8% 7.9% 16.1% 16.9% Table 11 Average improvement rates in comparison to the standardized aproach with a single profile attained by applying k = 2, 3, 4, 5, and 6 profiles for the quadratic inconvenience objective function (see Ferrucci and Bock (2016), p. 366) Quadratic Number of vehicles 8 10 12 Number of generated profiles Applied patterns 2p 3p 4p 2p 3p 4p 2p 3p 4p k=2 10.7% 14.2% 1.9% 11.4% 15.8% 4.0% 16.9% 19.7% 2.6% k=3 9.8% 20.6% 6.5% 11.5% 23.1% 15.4% 15.8% 31.8% 16.0% k=4 10.2% 19.8% 16.0% 8.6% 22.0% 23.9% 14.2% 31.1% 30.1% k=5 13.7% 17.8% 14.3% 12.3% 22.5% 23.7% 17.5% 31.6% 30.9% k=6 8.5% 15.3% 15.8% 12.2% 20.2% 25.6% 18.0% 29.4% 30.4% As in these settings some profiles contain outliers, a large value of Δmax t,p prevents the rejection of these profiles. This reduces the quality of the applied stochastic knowledge and leads to most significant improvements attained by using ER instead of SR. As ER determines the rejection of a profile by comparing it to the best matching profile, these outliers do not have a substantial effect (see Ferrucci and Bock (2016) for more detailed analyses of the progression of identified profiles done by SR and ER). Due to these results, the following tests are solely conducted by applying the pro-active real-time variant that uses ER. Tables 10 and 11 give the attained improvement rates of applying the pro-active-ER approach assuming k = 2, 3, 4, 5, and 6 profiles in comparison to the original single profile approach while applying the linear and the quadratic customer inconvenience function. The improvement rates attained for the two objective functions strongly indicate that the approach proposed by Ferrucci and Bock (2016) is able to identify existing Pro-Active Strategies in Online Routing 237 patterns in the past data to derive stochastic knowledge of higher quality. While the original approach is not able to derive suitable dummy requests from the past data due to interfering effects, the fast identification of a relevant subset of useful profiles done by the pro-active-ER approach leads to substantial reductions of the resulting request response times. This is particularly true if the number of assumed profiles k reaches the actual number of patterns. Increasing k to this value enables the proactive-ER approach to delete interferences by non-relevant days almost completely. Moreover, changes of the request arrival structure that may occur throughout the ongoing day are efficiently handled by updating the set of relevant profiles for each new time slice. As the attained improvement rates are quite stable when k is further increased, an efficient practical application of the approach does not require to exactly estimate the number of existing patterns. Rather, it is sufficient not to significantly underestimate this number while overestimation (within a reasonable range) does not significantly reduce the attained performance. This improves the practicability of the proposed approach. 8 Brief Summary This paper provides a brief overview of some recent pro-active real-time routing approaches that are designed for the delivery processes of urgent goods or services. These pro-active real-time approaches are characterized by the fact that stochastic knowledge about future request arrivals that is used for guiding vehicles into request-likely areas is derived from past data during a prior offline step. By combining time-spatial segments of past days, future request occurrences are forecasted. By doing so, no unverified assumptions are made concerning the availability of a specific distribution. This improves the practical applicability of the proposed approaches. In comparison to a reactive real-time approach that keeps each idle vehicle at its last service location, the pro-active approach of Ferrucci et al. (2013) guides vehicles to request-likely areas and keeps them there for a specific amount of time. It is shown that this approach leads to substantial improvements of the request response times if the request arrival data possess a minimum degree of diversity. By defining the so-called degree of structural diversity as a general measure, it is possible to reliably estimate the attainable performance of additionally using the derived stochastic knowledge. Subsequently, extensions of this pro-active approach are described. In order to efficiently control en route diversions that may increase the number of accidents occurring, the approach of Ferrucci and Bock (2015) integrates a customizable penalty cost value into the objective function. By doing so, additional en route diversions have to compensate the defined penalty costs with corresponding reductions of the request response times. It turns out that, depending on the pursued objectives, reasonable definitions of the penalty cost value can be derived. Finally, the pro-active approach can be extended to be able to identify multiple patterns in past request data. By doing so, the approach is capable of filtering out past days that are relevant for the current ongoing day in real time. As these selections avoid 238 S. Bock the unwanted interfering effects of past days that are irrelevant to the current day, the quality of the derived stochastic knowledge is frequently significantly improved. This may lead to further substantial reductions of request response times. Acknowledgments I thank DDS Digital Data Services GmbH in Karlsruhe, Germany, for generously providing us with excellent real-world road network data. References Arthur, D., & Vassilvitskii, S. (2007). k-means++: The advantages of careful seeding. In Proceedings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms, pp. 1027–1035 (PDF). Bock, S. (2010). Real-time control of freight forwarder transportation networks by integrating multimodal transports and multiple transshipments. European Journal of Operational Research, 200(3), 733–746. https://doi.org/10.1016/j.ejor.2009.01.046. Bock, S., Rosenberg, O., & van Brackel, T. (2006). Controlling mixed-model assembly lines in real-time by using distributed systems. European Journal of Operational Research, 168(3), 880–904. https://doi.org/10.1016/j.ejor.2004.07.035. Dantzig, G., & Ramser, J. (1959). The truck dispatching problem. Management Science, 6(1), 80– 91. https://doi.org/10.1287/mnsc.6.1.80. Davis, M. M., & Maggard, M. J. (1990). An analysis of customer satisfaction with waiting times in a two-stage service process. Journal of Operations Management, 9(3), 324–334. https://doi.org/ 10.1016/0272-6963(90)90158-A. Dijkstra, E. W. (1959). A note on two problems in connexion with graphs. Numerische Mathematik, 1, 269–271. https://doi.org/10.1007/BF01386390. Ferrucci, F. (2013). Pro-active dynamic vehicle routing: Real-time control and requestforecasting – Approaches to improve customer service. PhD dissertation at the University of Wuppertal. Springer Science&Business Media. https://doi.org/10.1007/978-3-642-33472-6, ISBN: 978-3-642-33472-6. Ferrucci, F., & Bock, S. (2014). Real-time control of express pickup and delivery processes in a dynamic environment. Transportation Research Part B Methodological, 63, 1–14. https:// doi.org/10.1016/j.trb.2014.02.001. Ferrucci, F., & Bock, S. (2015). A general approach for controlling vehicle en-route diversions in dynamic vehicle routing problems. Transportation Research Part B Methodological, 77, 76–87. https://doi.org/10.1016/j.trb.2015.03.003. Ferrucci, F., & Bock, S. (2016). Pro-active real-time routing in applications with multiple request patterns. European Journal of Operational Research, 253(2), 356–371. https://doi.org/10.1016/ j.ejor.2016.02.016. Ferrucci, F., Bock, S., & Gendreau, M. (2013). A proactive real-time control approach for dynamic vehicle routing problems dealing with the delivery of urgent goods. European Journal of Operational Research, 225(3), 130–141. https://doi.org/10.1016/j.ejor.2012.09.016. Gendreau, M., Guertin, F., Potvin, J.-Y., & Taillard, E. (1999). Parallel Tabu search for realtime vehicle routing and dispatching. Transportation Science, 33(4), 381–390. https://doi.org/ 10.1287/trsc.33.4.381. Giaglis, G. M., Minis, I., Tatarakis, A., & Zeimpekis, V. (2004). Minimizing logistics risk through real-time vehicle routing and mobile technologies. International Journal of Physical Distribution & Logistics Management, 34(9), 749–764. https://doi.org/10.1108/09600030410567504. Ichoua, S., Gendreau, M., & Potvin, J.-Y. (2000). Diversion issues in real-time vehicle dispatching. Transportation Science, 34(4), 426–438. https://doi.org/10.1287/trsc.34.4.426.12325. Pro-Active Strategies in Online Routing 239 Ichoua, S., Gendreau, M., & Potvin, J.-Y. (2006). Exploiting knowledge about future demands for real-time vehicle dispatching. Transportation Science, 40(2), 211–225. https://doi.org/10.1287/ trsc.1050.0114. Israel, D. (2008). Data analysis in business research: A step-by-step nonparametric approach. Los Angeles, CA: Response Books SAGE, ISBN: 9788178298757. Jain, A. K., Murty, M. N., & Flynn, P. J. (1999). Data clustering: A review. ACM Computing Surveys, 31(3), 264–323. https://doi.org/10.1145/331499.331504. Kaber, D. B., Liang, Y., Zhang, Y., Rogers, M. L., & Gangakhedkar, S. (2012). Driver performance effects of simultaneous visual and cognitive distraction and adaptation behavior. Transportation Research Part F, 15, 491–501. https://doi.org/10.1016/j.trf.2012.05.004. Kristensen, K., Kanji, G. K., & Dahlgaard, J. J. (1992). On measurement of customer satisfaction. Total Quality Management & Business Excellence, 3(2), 123–128. https://doi.org/10.1080/ 09544129200000013. Kvam, P. H., & Vidakovic, B. (2007). Nonparametric statistics with applications to science and engineering (Wiley series in probability and statistics). Hoboken, NJ: Wiley, ISBN: 0470081473. Larsen, A. (2000). The dynamic vehicle routing problem. PhD thesis at the Technical University of Denmark. IMM-PHD, No. 2000-73. Larsen, A., Madsen, O. B. G., & Solomon, M. M. (2008). Recent developments in dynamic vehicle routing systems. In B. L. Golden, S. Raghavan, & E. A. Wasil (Eds.), The vehicle routing problem: Latest advances and new challenges (Operations research/Computer science interfaces series) (Vol. 43, pp. 199–218). Heidelberg: Springer. https://doi.org/10.1007/978-0387-77778-8_9. Lund, K., Madsen, O. B. G., & Rygaard, J. M. (1996). Vehicle routing problems with varying degrees of dynamism. Technical report 1/96 at the Institute of Mathematical Modelling (IMM). Owens, J. M., McLaughlin, S. B., & Sudweeks, J. (2010). Driver performance while text messaging using handheld and in-vehicle systems. Accident Analysis & Prevention, 43, 939–947. https:// doi.org/10.1016/j.aap.2010.11.019. Park, S. K., & Miller, K. W. (1988). Random number generators: Good ones are hard to find. Communications of the ACM, 31(10), 1192–1201. https://doi.org/10.1145/63039.63042. Park, S. K., Miller, K. W., & Stockmeyer, P. K. (1993). Technical correspondence: Remarks on choosing and implementing random number generators. Communications of the ACM, 36(7), 105–110. https://doi.org/10.1145/159544.376068. Pillac, V., Gendreau, M., Guéret, C., & Medaglia, A. L. (2013). A review of dynamic vehicle routing problems. European Journal of Operational Research, 225(1), 1–11. https://doi.org/ 10.1016/j.ejor.2012.08.015. Rodgers, J. L., & Nicewander, W. A. (1988). Thirteen ways to look at the correlation coefficient. The American Statistician, 42(1), 59–66. https://doi.org/10.2307/2685263. Ross, S. M. (2019). Introduction to probability models (12th ed.). Amsterdam, Boston: Academic. ISBN: 9780128143469. Sapsford, R., & Jupp, V. (2006). Data collection and analysis (2nd ed.). London: Sage, 9780761943631. Young, K. L., Salmon, P. M., & Cornelissen, M. (2013). Distraction-induced driving error: An onroad examination of the errors made by distracted and undistracted drivers. Accident Analysis & Prevention, 58, 218–225. https://doi.org/10.1016/j.aap.2012.06.001. Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces M. Vimala Rani and M. Mathirajan 1 Introduction Semiconductors are required by many industries such as information technology (IT), IT enabled services, office automation, industrial machinery and automation, and engineering. Thus, the sales of semiconductors keep on increasing (SGSR 2020), making the semiconductors industry a crucial player in the industry sectors. Semiconductor manufacturing (SM) processes are broadly classified into the front-end process and back-end process. Wafer fabrication (Wafer fab) and wafer probing together constitute the front-end process. The assembly and final testing constitute the back-end process. In SM industry, supply chain management (SCM) problems are very complex, and this is due to long cycle times, high levels of stochasticity, and non-linearity in the manufacturing process (Sun and Rose 2015; Mönch et al. 2018). Further, in SM industry, the capability of meeting delivery commitment given to the customer as well as minimizing the cycle time are the most important issues in meeting the competition in the global market (Mönch et al. 2013). One of the ways to address this supply chain issue is to propose optimal/efficient scheduling, and this involves (i) maximizing the resource(s) usage, (ii) delivering the product on-time, and (iii) facing the increased demand and addressing the continuous competition in the marketplace. Researches in scheduling have addressed different areas of the SM industry (Koo and Moon 2013; Monch M. Vimala Rani () Vinod Gupta School of Management, Indian Institute of Technology, Kharagpur, Kharagpur, India e-mail: vimala@vgsom.iitkgp.ac.in M. Mathirajan Department of Management Studies, Indian Institute of Science, Bangalore, India e-mail: msdmathi@iisc.ac.in © Springer Nature Switzerland AG 2021 S. Srinivas et al. (eds.), Supply Chain Management in Manufacturing and Service Systems, International Series in Operations Research & Management Science 304, https://doi.org/10.1007/978-3-030-69265-0_9 241 242 M. Vimala Rani and M. Mathirajan et al. 2011; Mathirajan and Sivakumar 2006). This study addresses a scheduling problem in the manufacturing of wafer fab, which is the first phase process in SM. Scheduling is very important in wafer fab due to complex operations involving multiple types of expensive machines, re-entrant systems, and time-consuming processes. Thus, this study is concerned about scheduling in wafer fab, particularly diffusion operation. The diffusion operation is carried out in a diffusion furnace (DF), which is the batch processor (BP). In BP, more than one job is processed simultaneously (with the same start and finish times) as a batch subject to the capacity of the BP. The efficient scheduling of diffusion furnaces would significantly improve the performance of the SM industry’s supply chain, as the diffusion process is the lengthiest process (9–10 h) in the wafer fab area. Because of the long processing time required for the diffusion operation, the production rate, as well as due-date compliance of the product, are highly dependent on the efficient way of scheduling DF(s). Further, in reality, there are real-time events (RTE) dynamically occurring associated with job(s) (such as change in due-date, change in arrival time, change in priority, job cancellation) and/or resource(s) (e.g., machine breakdown, operator illness, tool failure, shortage of material, defective material) (Wang and Fei 2014). Due to this, in general, the production managers are finding difficulty in meeting the commitment given to the customer for delivering the product with the existing schedule. So, the production managers not only need to generate high-quality solutions, but they also have to address the real-time dynamic event in a costeffective manner. Hence, this study focuses on dynamic real-time scheduling (DRTS) of diffusion furnace(s). Further, as the processing time required at the diffusion furnace accounts for nearly 75% of the processing time required for the wafer fab, we considered different due-date based objectives. The organization of the chapter is as follows. The description of the research problem is given in the next section. Section 3 reviews the relevant previous research done in the scheduling of diffusion furnaces. In Sect. 4, the proposed mathematical model is discussed. Seven different Apparent Tardiness Cost (ATC) based greedy heuristic algorithms (GHA) are proposed and the same are discussed in Sect. 5. Analyses of the proposed seven different ATC-GHA are presented in Sect. 6. In the final section, the conclusion and potential future research directions are discussed. 2 Problem Description and Assumption The diffusion furnace available for scheduling may be a single diffusion furnace, identical parallel diffusion furnaces, or non-identical parallel diffusion furnaces. Further, due to technical reasons, some jobs can be processed only in a particular diffusion furnace available in the shop floor (this refers to machine eligibility restriction in scheduling). This study focuses on both single diffusion furnace and non-identical parallel diffusion furnaces with machine eligibility restrictions. The diffusion furnace simultaneously processes 6 to 12 standard jobs as a batch. Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 243 There are ‘N’ jobs available. Each job belongs to a different family, and each family has a different processing time. However, the jobs having the same processing time requirement belong to the same family. Due to the distinct nature of the chemical process in each family (incompatible-job families), a batch can be formed by considering a set of jobs from the same family. Further, jobs are having different arrival time, different due-date, and they are non-agreeable. That is, if the arrival time of job ‘i’ is less than the arrival time of job ‘j’ then it is not-implied that due date of job ‘i’ is less than the due date of job ‘j’. Further, the random occurrences of real time events with respect to job(s)/resources(s) while scheduling diffusion furnace(s) are considered. With this problem description, the objective of this study is to optimize eight different customer perspectives objectives/criteria: Total Tardiness (TT), Total Weighted Tardiness (TWT), Number of Tardy (NT) Jobs, Weighted NT (WNT) Jobs, On Time Delivery (OTD) rate, Total Earliness and lateness (TEL), Total Weighted Earliness and Lateness (TWEL), and Maximum Lateness (Lmax). These criteria are defined as follows: N TT = Tj j=1 TWT = N Wj ∗ Tj j=1 NT = N UPj j=1 WNT = N Wj ∗ UPj j=1 OTD Rate = N OTDj/N j=1 N CTj − DDj TEL = j=1 TWEL = N Wj ∗ CTj − DDj j=1 Lmax = max Lj. j =1 t o N 244 M. Vimala Rani and M. Mathirajan Where, CTj : DDj : Tj : Wj : UPj : OTDj : Lj : Completion time of job ‘j’ Due-date of job ‘j’ Tardiness of job ‘j’ = (CTj - DDj ) if CTj > DDj ; 0 Otherwise Weight of job ‘j’ Unit penalty of job ‘j’ = 1 if CTj > DDj ; 0 Otherwise On time delivery of job ‘j’ = 0 if CTj > DDj ; 1 Otherwise Lateness of job ‘j’ = (CTj - DDj ) This study makes the following assumptions for developing prescriptive analytics model(s)/methodologies: • Other than RTE, which occurs randomly during the schedule planning period, we assumed the deterministic situation. • Every job must pass through the diffusion operation, and it is not dependent on other jobs • For each diffusion furnace, the number of jobs in a batch should be less than or equal to the maximum limit of the corresponding diffusion furnace. • Preemption is not allowed 3 Related Literature Review Among the four phases of SM, wafer fab is the highly complex and capitalintensive phase (Li et al. 2017). The importance of scheduling in wafer fab has been increasing steadily over the past few decades (Sarin and Shikalgar 2001). This area takes a total of 3–15 weeks in comparison with the required overall processing time of 8–30 weeks for SM. The longer processing times often belong to batch processing operations (diffusion, oxidation, deposition, etc.,) in wafer fab. Scheduling of batch processing machines causes higher machine utilization, lower work-in-process inventory, shorter cycle time, and greater customer satisfaction (Pinedo 2012). This study particularly focuses on the scheduling of diffusion furnaces, a batch processor, where the diffusion operation takes up to 10 h (Mönch et al. 2006). Furthermore, around 30% of the total WIP (Work in Process) in a wafer fab lie in diffusion stages, due to long process times (Jung et al. 2014). Based on the type of data, diffusion furnace(s) scheduling can be grouped into deterministic and stochastic. This study focuses on the deterministic scheduling of diffusion furnace(s) as data tracking is not a difficult one in the SM industry, where the whole shop floor is computerized. However, readers interested in stochastic scheduling of BPM can refer to Park and Banerjee (2011). Some of the researchers (e.g. Pirovano et al. 2020) in deterministic scheduling of diffusion furnace(s) consider one or more upstream and/or downstream operations along with the diffusion operation. This study concerns about single bottleneck operation: diffusion operation only. The diffusion furnace(s) considered for research Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 245 in wafer fab may be grouped into single diffusion furnace, identical parallel diffusion furnaces, non-identical parallel diffusion furnaces, diffusion furnaces with machine eligibility restrictions, and diffusion furnaces without machine eligibility restrictions. This study focuses on both single and non-identical parallel diffusion furnaces with machine eligibility restrictions. Further, diffusion furnace(s) scheduling can be grouped into static scheduling (e.g. Mönch and Roob 2018) and dynamic scheduling. Various DF studies addressing the dynamic scheduling of DF(s) consider the future arrival of jobs related to the dynamic situation. However, in reality, there is another situation related to the dynamic situation, i.e., RTE. Hence, this study considers both the dynamic situations: dynamic job arrival and the occurrences of RTE randomly. Based on objectives, the studies on deterministic and dynamic scheduling of diffusion furnace(s) are further classified into (i) completion time-based objectives (related to organization perspective), and (ii) due-date based objectives (related to customer perspective). Because the present study considers customer perspectives objective, the organizational perspectives objectives (completion time-based objectives) (e.g. Rocholl et al. 2018) are not discussed here. Finally, based on the solution methodologies, the existing research can be broadly classified into mathematical programming-based approaches, greedy heuristic approaches, metaheuristic approaches, and simulation. Based on the classifications presented here, a brief review of closely related earlier studies are summarized in Table 1. Table 1 clearly indicates that the research problem configuration considered (last row of Table 1) in this study: “dynamic real-time scheduling of (a) single diffusion furnace, and (b) non-identical parallel diffusion furnaces with machine eligibility restrictions, dynamic jobarrivals, incompatible-job families, non-agreeable release times & due-dates, and unexpected RTE, to optimize eight different due-date based scheduling objectives” is a new scheduling problem associated with batch processor in general, particularly diffusion furnaces. 4 Mathematical Model for Dynamic Scheduling of Diffusion Furnaces The proposed mathematical models for dynamic scheduling (DS) of single diffusion furnace (SDF) and non-identical parallel diffusion furnaces with machine eligibility restrictions (NPDF-MER) to optimize eight different objectives/criteria are explained in this section. In the proposed mathematical models, only the future arrival of the jobs is considered under dynamic scheduling, whereas the randomly occurring real-time events and dynamic arrival of jobs are considered in the proposed ATC based greedy heuristic algorithms. Further, this section discusses the validation of the proposed mathematical model. Kurz and Mason (2008) Chiang et al. (2008) Cheng et al. (2008) Malve and Uzsoy (2007) Mönch et al. 2006 Dirk and Monch (2006) Mönch et al. (2005) Kurz (2003) Uzsoy (1995) Author √ √ √ √ √ √ √ √ √ √ √ √ √ √ With machine Eligibility √ Without machine Eligibility √ With machine Eligibility Non-identical Parallel machine With future arrival of jobs With With job resource related related RTE RTE Only Only Dynamic Scheduling √ Without machine Single Eligibilmachine ity √ Identical Parallel Machine Diffusion furnace as Table 1 A closely related review √ √ √ √ √ √ √ √ √ √ Mathematical Programming √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ (continued) √ √ √ OTD TT TWT NT WNT rate TEL TWEL Lmax Scheduling objective Wt. of Greedy Meta onHeuris- Heuris- time tic tic jobs Release time & due-date Solution methodologies With Job and resource Nonrelated Agreeable agreeable RTE Vimala Rani and Mathirajan (2016b) Vimala Rani and Mathirajan (2016a) Vimala Rani and Mathirajan (2015) Mansoer and Koo (2015) Bilyk et al. (2014) Jia et al. (2013b) Chen et al. (2013) Jia et al. (2013a) Mathirajan and Vimalarani (2012) Li et al. (2010) Kim et al. (2010) Guo et al. (2010) Chiang et al. (2010) Bar-Noy et al. (2009) Li et al. (2009) Li and Qiao (2008) Table 1 (continued) √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ (continued) √ Proposed research study Vimala Rani and Mathirajan (2020b) Vimala Rani and Mathirajan (2020a) Fidelis and Arroyo (2017) Vimala Rani and Mathirajan (2016c) Author √ √ √ √ √ √ With machine Eligibility √ Without machine Eligibility √ With machine Eligibility Non-identical Parallel machine With future arrival of jobs With With job resource related related RTE RTE Only Only Dynamic Scheduling √ √ Without machine EligibilSingle machine ity √ Identical Parallel Machine Diffusion furnace as Table 1 (continued) √ √ √ √ √ √ √ √ √ √ √ √ √ Mathematical Programming √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ √ OTD TT TWT NT WNT rate TEL TWEL Lmax Scheduling objective Wt. of onGreedy Meta Heuris- Heuris- time jobs tic tic Release time & due-date Solution methodologies With Job and resource related NonRTE Agreeable agreeable Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 249 4.1 (0-1) MILP Model for DS-SDF The notations used for the proposed mathematical model for DS-SDF are discussed as follows. Notations: Sets: J Jobs F Families B Batches Index: j 1 . . . N for jobs b 1 . . . K for batches f 1 . . . G for families Parameters: A First time availability of the DF BC Batch capacity of DF RTj Release time of a job ‘j’ DDj Due-date of a job ‘j’ Wj Priority (or) Weight of a job ‘j’ PTf Processing time of a family ‘f’ FAjf Family association for a job, equals 1 if job ‘j’ belong to family ‘f’; 0 otherwise Decision variables: Xjb Yfb 1 if job ‘j’ is processed in a batch ‘b’; 0 Otherwise 1 if family ‘f’ is processed in a batch ‘b’; 0 Otherwise Dependent variables: CTBb Completion time of a batch ‘b’ RTBb Release time of a batch ‘b’ PTBb Processing time of a batch ‘b’ CTj Completion time of job ‘j’ Tj Tardiness of job ‘j’ ELj Earliness lateness of job ‘j’ UPj Unit penalty of job ‘j’ TT Total tardiness of ‘N’ jobs TWT Total weighted tardiness of ‘N’ jobs NT Number of tardy jobs among ‘N’ jobs WNT Weighted number of tardy jobs OTDRate On time delivery rate among ‘N’ jobs TEL Total earliness lateness among ‘N’ jobs 250 M. Vimala Rani and M. Mathirajan TWEL Total weighted earliness lateness of ‘N’ jobs Lmax Maximum lateness among ‘N’ jobs Formulation of (0-1) MILP Model for DS-SDF Minimize T T (1) Subject to G Yf b ≤ 1 ∀b ∈ [1, K] (2) ∀b ∈ [2, K] (3) f =1 G Yf,b−1 ≥ f =1 G Yf b f =1 Xj b ∗ FAjf ≤ Yf b ∀j ∈ [1, N] ; ∀b ∈ [1, K] ; ∀f ∈ [1, G] (4) N Xj b ∗ FAjf ≥ Yf b ∀f ∈ [1, G] ; ∀b ∈ [1, K] (5) Xj b = 1 ∀j ∈ [1, N] (6) Xj b ≤ BC ∀b ∈ [1, K] (7) RTB b ≥ A b=1 (8) RTB b ≥ RTj ∗ Xj b ∀j ∈ [1, N] ; ∀b ∈ [1, K] (9) RTB b ≥ CTB b−1 ∀b ∈ [2, K] (10) ∀b ∈ [1, K] (11) ∀b ∈ [1, K] (12) j =1 K b=1 N j =1 P TB b = G PTf ∗ Yf b f =1 CTB b = RTB b + P TB b Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces CTj ≥ CTB b − (BigM) ∗ 1 − Xjb 251 ∀b ∈ [1, K] ; ∀j ∈ [1, N] (13) Xj b ∈ {0, 1} ∀j ∈ [1, N] ; ∀b ∈ [1, K] (14) Yf b ∈ {0, 1} ∀f ∈ [1, G] ; ∀b ∈ [1, K] (15) Computation of Due-date based Scheduling Criteria Tj ≥ CTj − DD j TT = N J =1 T WT = N ∀j ∈ [1, N] (16) (17) Tj Wj ∗ Tj (18) j =1 UPj ≥ min 1, Tj NT = N ∀j ∈ [1, N] UPj (19) (20) j =1 N W NT = Wj ∗ UPj j =1 ⎛ OT D Rat e = ⎝N − N (21) ⎞ UPj ⎠ /N (22) j =1 ELj ≥ CT j − DD j T EL = N ELj ∀j ∈ [1, N] (23) (24) j =1 T W EL = N Wj ∗ ELj (25) j =1 Lmax = max Tj j =1 t o N (26) 252 M. Vimala Rani and M. Mathirajan The objective function (Eq. 1) seeks to minimize the total tardiness of ‘N’ jobs. Constraint (2) ensures that no more than one family is assigned to a batch. Constraint (3) takes a responsibility to construct a batch sequentially. Constraint (4) states that a job can be processed in a batch only if the corresponding family is processed in that batch. Constraint (5) conditions that empty family (that is, a family without any jobs) cannot be assigned to any batch. Constraint (6) is used to assign the job to any one batch without splitting it. Constraint (7) ensures that the number of jobs in any batch does not exceed the batch capacity of the diffusion furnace. Constraint (8) guarantees that the first batch can start after the first-time availability (that is the given, ‘Ath’ hour) of the diffusion furnace. Moreover, the starting time of every batch should be greater than the release time of jobs in that batch and the completion time of the previous batch. Constraint (9) and (10) ensure this. Constraint (11) states that the processing time of a batch is equal to the processing time of the corresponding family. Constraint (12) computes the completion time of each batch by adding its starting time and processing time. Constraint (13) computes the completion time of the job. Constraint (14) and Constraint (15) assign binary values to decision variables. Constraint (16) and (17) are used to compute tardiness and total tardiness, respectively. For primary scheduling objective represented in constraint (1), Constraints (18) to (26) are introduced to compute the value of other scheduling objectives/criteria. Accordingly, Constraint (18) calculates total weighted tardiness. Constraint (19) assigns ‘1’ to jobs that are completed after their due-date. Constraint (20) and (21) calculate the total number of tardy jobs and the total weight of all tardy jobs, respectively. Constraint (22) is used to calculate on time delivery rate. A Constraint (23) calculates the earliness lateness as the absolute difference between the completion time of a job and its due-date. Constraint (24) calculates the total earliness lateness, and Constraint (25) calculates the total weighted earliness lateness of the job. Finally, Constraint (26) calculates maximum lateness among ‘N’ jobs. Like minimizing total tardiness (TT) of ‘N’ jobs as a primary scheduling objective, we can optimize any one the following scheduling criterion/objective as a primary scheduling criterion/objective as follows: 1. To minimize TWT, we need to have the Constraints (2) to (16) and Constraint (18) along with Constraint (1) as a new objective function: Min TWT 2. To minimize NT, we need to have the Constraints (2) to (16), Constraint (19), and Constraint (20) along with Constraint (1) as a new objective function: Min NT 3. To minimize WNT, we need to have the Constraints (2) to (16), Constraint (19), and Constraint (21) along with Constraint (1) as a new objective function: Min WNT 4. To maximize OTD rate, we need to have the Constraints (2) to (16), Constraint (19) and Constraint (22) along with Constraint (1) as a new objective function: Max OTD rate Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 253 Table 2 A numerical instance Job J1 J2 J3 J4 J5 J6 J7 J8 J9 J10 Family 2 1 4 3 5 1 1 3 1 5 Release time 1 6 4 1 4 8 6 2 6 6 Due-date 26 31 23 16 31 29 28 30 36 26 Priority (or weight) 6 2 3 6 3 2 6 1 2 1 5. To minimize TEL, we need to have the Constraints (2) to (15), Constraint (23), and Constraint (24) along with Constraint (1) as the new objective function: Min TEL 6. To minimize TWEL, we need to have the Constraints (2) to (15), Constraint (23), and Constraint (25) along with Constraint (1) as the new objective function: Min TWEL 7. To minimize Lmax, we need to have the Constraints (2) to (16) and Constraint (26) along with Constraint (1) as the new objective function: Min Lmax Validation of the Proposed (0-1) MILP Model for DS-SDF: To validate the proposed (0-1) MILP model, numerical data (Table 2) representing the research problem is generated. Further, this study assumed that maximum of two jobs could be assigned to a batch, and the furnace is available from the second hour onwards. This numerical example is given as input to the LINGO set code for generating the proposed (0-1) MILP model and solved using the LINGO solver. The optimal solution obtained is presented in Table 3. The analysis of this table confirms the right mathematical representation of the DS-SDF. 4.2 (0-1) MILP Model for DS-NPDF-MER The additional/modified notations required towards the proposed model for DSNPDF-MER are given here. Additional Notations Sets: M Machines (Diffusion furnaces) Index: m 1 . . . L for machines Parameters: 254 M. Vimala Rani and M. Mathirajan Table 3 The optimal solution Batch B1 B2 Job Batch (B) RT PT CT 2 4 6 6 10 16 J1 J4 J8 B3 J2 16 2 J6 B4 J7 18 2 J9 B5 J5 20 20 J10 B6 J3 40 16 Total tardiness B1 J4 2 10 J8 B2 J2 12 2 J7 B3 J6 14 2 J9 B4 J1 16 4 B5 J5 20 20 J10 B6 J3 40 16 Total weighted tardiness B1 J1 2 4 B2 J4 6 10 B3 J8 16 10 B4 J6 26 2 J7 B5 J2 28 2 J9 B6 J5 30 20 J10 B7 J3 50 16 Number of tardy jobs B1 J4 3 10 B2 J6 13 2 J7 B3 J1 15 4 B4 J8 19 10 B5 J2 29 2 J9 18 20 40 56 12 14 16 20 40 56 6 16 26 28 30 50 66 13 15 19 29 31 Job DD 26 16 30 31 29 28 36 31 26 23 W 6 6 1 2 2 6 2 3 1 3 16 30 31 28 29 36 26 31 26 23 6 1 2 6 2 2 6 3 1 3 26 16 30 29 28 31 36 31 26 23 6 6 1 2 6 2 2 3 1 3 16 29 28 26 30 31 36 6 2 6 6 1 2 2 T 0 0 0 0 0 0 0 9 14 33 56 0 0 0 0 0 0 0 9 14 33 E L WT UPJ WNT EL WEL 0 0 0 0 0 0 0 27 14 99 140 0 0 0 0 0 0 0 1 1 1 3 0 0 0 0 0 0 0 (continued) Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 255 Table 3 (continued) Batch (B) Job RT PT CT DD W T B6 J5 31 20 51 31 3 B7 J3 51 16 67 23 3 B8 J10 67 20 87 26 1 Weighted number of tardy jobs B1 J1 2 4 6 26 6 B2 J4 6 10 16 16 6 B3 J8 16 10 26 30 1 B4 J6 26 2 28 29 2 J7 28 6 B5 J2 28 2 30 31 2 J9 36 2 B6 J5 30 20 50 31 3 J10 26 1 B7 J3 50 16 66 23 3 On time delivery rate (10- unit penalty)/10 B1 J4 2 10 12 16 6 J8 30 1 B2 J5 12 20 32 31 3 J10 26 1 B3 J2 32 2 34 31 2 J6 29 2 B4 J7 34 2 36 28 6 J9 36 2 B5 J1 36 4 40 26 6 B6 J3 40 16 56 23 3 Total earliness lateness B1 J4 2 10 12 16 6 J8 30 1 B2 J3 12 16 28 23 3 B3 J6 28 2 30 29 2 J7 28 6 B4 J1 30 4 34 26 6 B5 J2 34 2 36 31 2 J9 36 2 B6 J5 36 20 56 31 3 J10 26 1 Total weighted earliness lateness B1 J1 2 4 6 26 6 B2 J4 6 10 16 16 6 J8 30 1 B3 J3 16 16 32 23 3 Batch Job E L WT UPJ WNT 3 3 1 7 EL WEL 0 0 0 0 0 0 0 1 1 1 0.7 4 18 0 0 0 0 0 0 0 0 0 0 1 6 3 5 8 0 14 33 4 18 0 0 0 0 0 0 0 0 0 0 5 1 2 8 5 0 25 30 4 18 1 6 3 5 8 0 14 33 92 24 18 15 2 12 48 10 0 75 30 234 0 0 0 9 (continued) 256 M. Vimala Rani and M. Mathirajan Table 3 (continued) Batch Job B4 Batch (B) RT PT CT 32 20 52 J5 J10 B5 J6 52 2 J7 B6 J2 54 2 J9 Maximum lateness 54 56 Job DD 31 26 29 28 31 36 W 3 1 2 6 2 2 T E L 21 26 25 26 25 20 26 WT UPJ WNT EL WEL Am First time availability of the machine ‘m’ BCm Batch capacity of machine ‘m’ MAjm Machine association for a job, equals 1 if job ‘j’ can be processed in machine ‘m’; 0 otherwise Decision variables: Xjbm Yfbm 1 if job ‘j’ is processed in a batch ‘b’ in a machine ‘m’; 0 Otherwise 1 if family ‘f’ is processed in a batch ‘b’ in a machine ‘m’; 0 Otherwise Dependent variables: RTBbm PTBbm CTBbm Release time of a batch ‘b’ in a machine ‘m’ Processing time of a batch ‘b’ in a machine ‘m’ Completion time of a batch ‘b’ in a machine ‘m’ Formulation of (0-1) MILP Model for DS-NPDF-MER Concept wise all the constraints in the proposed model for DS-SDF are like that of the model required for DS-NPDF-MER, except the Constraint (6) presented in Sect. 4.1. Furthermore, the constraints required to compute the value of TT, TWT, NT, WNT, OTD rate, TEL, TWEL, and Lmax are independent of the number of furnaces. Hence, the constraints presented in Sect. 4.1 for computing scheduling criteria/objectives are exactly the same for DS-NPDF-MER. Accordingly, the proposed model with the scheduling objective of minimizing TT for DS-NPDF-MER is as follows: Minimize T T Subject to G f =1 Yf bm ≤ 1 (27) ∀b ∈ [1, K] ; ∀m ∈ [1, L] (28) Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces G Yf,b−1,m ≥ f =1 G Yf bm ∀b ∈ [2, K] ; ∀m ∈ [1, L] 257 (29) f =1 Xj bm ∗ F Ajf ≤ Yf bm ∀j ∈ [1, N ] ; ∀f ∈ [1, G] ; ∀b ∈ [1, K] ; ∀m ∈ [1, L] (30) N Xj bm ∗ FAjf ≥ Yf bm ∀m ∈ [1, L] ; ∀f ∈ [1, G] ; ∀b ∈ [1, K] j =1 (31) K L Xj bm ∗ MAj m = 1 ∀j ∈ [1, N] (32) ∀b ∈ [1, K] ; ∀m ∈ [1, L] (33) RT B bm ≥ Am b = 1; ∀m ∈ [1, L] (34) RT B bm ≥ RT j ∗ Xj bm ∀b ∈ [1, K] ; ∀m ∈ [1, L] ; ∀j ∈ [1, N] b=1 N m=1 Xj bm ≤ BC m j =1 (35) RT B bm ≥ CT B b−1, P T B bm = G m P T f ∗ Yf bm ∀b ∈ [2, K] ; ∀m ∈ [1, L] (36) ∀b ∈ [1, K] ; ∀m ∈ [1, L] (37) ∀b ∈ [1, K] ; ∀m ∈ [1, L] (38) f =1 CT B bm = RT B bm + P T B bm CT j ≥ CT B bm − (BigM) ∗ 1 − Xj bm ∀j ∈ [1, N] ; ∀b ∈ [1, K] ; ∀m ∈ [1, L] (39) Xj bm ∈ {0, 1} ∀j ∈ [1, N] ; ∀b ∈ [1, K] ; ∀m ∈ [1, L] (40) Yf bm ∈ {0, 1} ∀f ∈ [1, G] ; ∀b ∈ [1, K] ; ∀m ∈ [1, L] (41) Tj ≥ CT j − DD j ∀j ∈ [1, N] (42) 258 M. Vimala Rani and M. Mathirajan Validation of the Proposed (0-1) MILP Model for DS-NPDF-MER: The same numerical problem presented in Table 2 is used to validate the model for DS-NPDFMER. In addition, three non-identical parallel diffusion furnaces: MC1, MC2 and MC3 are considered for the numerical example with capacity as 2 jobs, 3 jobs and 4 jobs, respectively. This study assumes that the first-time availability of these three non-identical parallel diffusion furnaces is to be 1st hour, 2nd hour, and 3rd hour, respectively. Further, jobs that belong to family 3 (f3) are processed only in MC2, out of three non-identical parallel diffusion furnaces considered here. The same numerical example, given in Table 2, along with required additional data presented here, is given as input to the LINGO set code for generating the proposed (0-1) MILP model and solving using the LINGO solver. The optimal solution reports, similar to Table 3, are prepared. Due to the brevity of this chapter, the reports are not presented here. However, by analyzing the reports prepared, the proposed model for DS-NPDF-MER has been validated. Since the mathematical model for DS-SDF to minimize TWT is computationally intractable (Vimala Rani and Mathirajan 2016a), the proposed mathematical model for DS-NPDF-MER is also computationally intractable. So, we focus on GHA based on dispatching rules, particularly Apparent Tardiness Cost (ATC) rule, and the same is presented in the next section. 5 ATC Based GHA for Scheduling Diffusion Furnaces Dispatching rules are widely used in the semiconductor manufacturing industry (Hildebrandt et al. 2010). The popularity of dispatching rule based GHA are derived from the fact that they (a) are efficient in a wide range of scheduling problems, (b) are generally easy to understand and computerize, (c) require very meagre computational time to provide solution, and (d) can deal with dynamic changes easily and quickly. Out of various due-date based dispatching rules, the ATC rule based GHA provides the most efficient solution (Vimala Rani and Mathirajan 2020a). Accordingly, this study also proposes the dispatching rule, ATC based GHA for DRTS of SDF as well as NPDF-MER, and the same is discussed in this section. 5.1 ATC-GHA for DRTS of SDF The details of ATC-GHA for DRTS of SDF are given below: Step 1: At the time of deciding a batch (batch selection) for scheduling a DF, (a) if any of the job(s) related RTEs occur, then modify the corresponding job-data on the work-in-process (WIP), and (b) if any of the resource(s) related RTEs occur Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 259 then modify the DF’s available time as T = T+ delay time due to resource related RTE. Step 2: At the time of deciding a batch for scheduling a DF, capture all the characteristics of the jobs which are waiting for diffusion operation. Further, capture the DF capacity (B) and the available time (T). Step 3: Cluster the jobs waiting in front of the DF based on their family. Step 4: In each cluster, calculate “Job-Priority-Index” for every job and sort the same. Step 5: For each cluster, create a “temporary batch” by picking ‘B’ jobs from the top. If the number of jobs in any of the temporary batch is not equal to ‘B’, then check the WIP for whether the jobs of the same family are coming for diffusion operation in the future. If it is true, then wait for those jobs to form a full batch. Otherwise, form a partially filled batch. Then modify the starting time of the temporary batch of that cluster as max (T, longest release time of all jobs in that temporary batch). Step 6: Compute and compare the starting time of each temporary batch, If any temporary batch has a completion time strictly less than the starting time of all other temporary batches, then select it and go to Step 9 Else go to the next step Step 7: Compute “Batch-Priority-Index” for each temporary batch formed in Step 5. Step 8: Choose the temporary batch that has the greatest “Batch-Priority-Index”. Step 9: If any of the job(s) related RTEs occur, then modify the corresponding jobdata on the work-in-process (WIP), else if any of the resource(s) related RTEs occur, then modify the DF’s available time as T = T+ delay time due to resource related RTE and go to Step 2. Else, go to the next step. Step 10: Assign the selected batch to the diffusion furnace. Step 11: Compute TWT, NT, WNT, OTD rate, TEL, TWEL, and Lmax of assigned batch (es). Step 12: Remove the assigned jobs from WIP and modify the DF’s available time “T” as the completion time of the assigned batch in Step 10 Step 13: Repeat Step 1 until all jobs are scheduled. This study proposed and implemented seven different above GHA in Turbo C. All these seven different GHA differ only in two steps (Step 4 and Step 7), and the same is presented in Table 4. 5.2 ATC-GHA for DRTS for NPDF-MER The details of ATC based GHA for DRTS of NPDF-MER are given below: Step 1: At the time of deciding a batch (batch selection) for scheduling a DF, (a) if any of the job(s) related RTEs occur, then modify the corresponding job-data on 260 M. Vimala Rani and M. Mathirajan Table 4 Modified steps in different ATC-GHA Proposed ATC-GHA ATC-GHA1 ATC-GHA2 ATC-GHA3 ATC-GHA4 ATC-GHA5 ATC-GHA6 ATC-GHA7 Step 6: Calculate Job-Priority-Index based on EDD ATC rule in Balasubramanian et al. (2004) ATC rule in Mönch et al. (2006) ATC rule in Li et al. (2010) ATC rule in Li et al. (2010) ATC rule in Vimala Rani and Mathirajan (2016a) ATC rule in Vimala Rani and Mathirajan (2016a) Step 9:Calculate Batch-Priority-Index based on Batch ATC (BATC) rule in Mehta and Uzsoy (1998) BATC rule in Balasubramanian et al. (2004) BATC rule in Mönch et al. (2006) BATC rule in Li et al. (2010) BATC rule in Mönch et al. (2006) BATC rule in Mönch et al. (2006) BATC rule in Vimala Rani and Mathirajan (2016a) the work-in-process (WIP), and (b) if any of the resource(s) related RTEs occur then modify the corresponding DF’s (say, DFi) available time as Ti = Ti + delay time due to resource related RTE. Step 2: At the time of deciding a batch for scheduling a DF, capture all the characteristics of the jobs, which are waiting for diffusion operation and the available time for each of the diffusion furnaces (say ti , ti > 0, where i = 1 to M and M indicates the number of diffusion furnace). Step 3: Select the diffusion furnace, which is available earlier. (i) If a tie occurs in terms of DF available time, then select the DF, which has the maximum capacity. (ii) If a tie occurs in terms of DF available time and maximum capacity, then select the DF, which has machine eligibility restriction. (iii) If a tie occurs in terms of DF available time, maximum capacity and machine eligibility restriction, then select the DF arbitrarily. Step 4: Store the selected DF available time and its capacity to ‘B’ & ‘T’ respectively. Step 5: Cluster the jobs waiting in front of the selected DF based on their family. Step 6: For each of the qualified cluster (that is, families, which are possible to process in a selected diffusion furnace), calculate “Job-Priority-Index” for every job and sort the same. Step 7: For each of the qualified clusters, create a “temporary batch” by picking ‘B’ jobs from the top. If the number of jobs in any of the temporary batch in any of the qualified clusters is not equal to ‘B’ then check the WIP for whether the jobs of the same family are coming for diffusion operation in the future. If it is true, then wait for those jobs to form a full batch. Otherwise, form a partially filled batch. Now, modify the starting time of the corresponding temporary batch as max (T, longest release time of all jobs in that temporary batch). Step 8: Compute and compare the starting time of each temporary batch, Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 261 If any temporary batch has a completion time strictly less than the starting time of all other temporary batches, then select it and go to Step 11 Else go to the next step Step 9: Compute Batch-Priority-Index for each temporary batch formed in Step 7. Step 10: Choose the temporary batch that has the greatest “Batch-Priority-Index”. Step 11: If any of the job(s) related RTEs occur, then modify the corresponding job-data on the work-in-process (WIP), and (b) if any of the resource(s) related RTEs occur, then modify the corresponding DF’s (say, DFi) available time as Ti = Ti + delay time due to resource related RTE and go to Step 2. Else, go to the next step. Step 12: Assign the selected batch to the selected diffusion furnace. Step 13: Compute TT, TWT, NT, WNT, OTD rate, TEL, TWEL, and Lmax of assigned batch(es) Step 14: Remove the assigned jobs from WIP and modify the selected DF’s available time “Ti ” as the completion time of the allocated batch in Step 12. Step 15: Repeat Step 1 until all jobs are scheduled. This study proposed seven different ATC-GHA for DRTS-NPDF-MER by varying Step 6 and Step 9, as per Table 4 and implement all these seven different ATC-GHA in Turbo C. 6 Performance Evaluation of ATC-GHA for DRTS of Diffusion Furnaces The performance of the seven different ATC-GHA for DRTS-SDF as well as DRTS-NPDF-MER are evaluated w.r.t. optimal solution on small scale data and estimated optimal solution (EOS) on large scale data. The EOS is computed based on Weibull distribution (Rardin and Uzsoy 2001). The details of the experimental design, performance measures, empirical and statistical analyses are presented in this section. Experimental Design The experimental design proposed in Vimala Rani and Mathirajan (2020a) is extended to represent the new research problem defined on scheduling SDF and NPDF-MER, with four additional factors: number of diffusion furnaces (M), capacity of diffusion furnace (B), available time of diffusion furnace (AT), and machine eligibility restriction (Mf ) to address non-identical parallel diffusion furnace with machine eligibility restrictions. Since this study compares the heuristics solution w.r.t optimal solution on small-scale data, the factor: number of jobs (N) has two more additional values such as 5 and 10. Table 5 shows the various factors and their values, which are used to generate 450 test data. Performance Measures for Empirical Analysis Average relative percentage deviation (ARPD) and Integrated Rank (IRANK) are used here. For each proposed 262 M. Vimala Rani and M. Mathirajan Table 5 An experimental design for DRTS of diffusion furnaces Parameters Number of diffusion furnaces (M) Capacity of diffusion furnace (B) Number of levels 1 1 Available time of diffusion furnace (AT) Number of jobs (N) Release time of jobs (RTj ) Due-date of jobs (DDj ) Family (f) Family processing time (PTf ) and the probability of job being in the family ‘f’ 1 Machine eligibility restriction (Mf ) Weight (Wj ) [low, high] No. of problem configurations No. of instances per configuration Total problem instances 5 3 3 1 1 1 Level wise values 4 [6, 6, 9, 12] for DF1 to DF4 respectively [2, 5, 7, 8] for DF1 to DF4 respectively 5, 10, 25,50,100 [1,8], [1,16], [1,24] [1,40], [1,60], [1,80] [1,5] [2,4,10,16,20] with a probability of [0.1, 0.3, 0.4, 0.1, 0.1] respectively f3 can be processed in only DF2 1 [1,10] 5*3*3*1*1*1 = 45 10 450 different ATC-GHA, this study computes deviation as per the Eq. (43), for each problem instances. To know the extent of this deviation w.r.t. benchmark solution (BSi ), this study divides the deviation of each different ATC-GHA for each problem instance (Dij ) by corresponding benchmark solution (BSi ) and converts it into percentage as per equation in (44). Dij = F S ij − BS i (43) RP D ij = Dij /BS i ∗ 100 (44) Where i: Problem instances and i ∈ [1,450] j: ATC-GHA and j ∈ [1,7] FSij : Feasible solution obtained from ‘jth ’ ATC-GHA for ‘ith ’ problem instance BSi : Benchmark solution for ‘ith ’ problem instance Dij : Deviation of ‘jth ’ ATC-GHA for ‘ith ’ problem instance RPDij : Relative percentage deviation of ‘jth ’ ATC-GHA for ‘ith ’ problem instance Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 263 Then, for each different ATC-GHA, the score on ARPD for each problem configuration as well as for 450 problem instances are computed as per Eq. (45) ARP D j = N i=1 RP D ij /N (45) Where ARPDj : Average relative percentage deviation of ‘jth ’ ATC-GHA N = 10 when ARPD is computed considering problem configuration wise N = 450 when ARPD is computed considering entire problem instances Further, for each of the different ATC-GHA, this study computes IRANK, which is proposed by Mathirajan et al. (2004), for each of the objectives considered in this study as per the Eq. (46). IRANKj = Maxrank r=1 Maxrank {N (r, j ) ∗ r} / N (r, j ) r=1 (46) Where r: Rank and r ∈ [1, 7] N (r, j): Number of times the ‘jth ’ ATC-GHA in rank ‘r’ Maxrank: Maximum Rank possible (Maxrank = 7 = Number of proposed ATCGHA). IRANKj : Integrated rank of ‘jth ’ ATC-GHA Performance Measures for Statistical Analysis For performance evaluation using statistical analyses, first we compute descriptive statistics: mean, median, standard deviation, and 95% confidence interval. As normality test is failed over the obtained RPD results, Kruskal-Wallis non-parametric test are conducted to compare the proposed different ATC-GHA (Beldar and Costa 2018). 6.1 Empirical Analyses on the Performance of ATC-GHA for DRTS of Diffusion Furnaces Empirical analyses of seven different ATC-GHA for DRTS of (i) SDF, and (ii) NPDF-MER w.r.t. the objectives: TT, TWT, NT, WNT, OTD rate, TEL, TWEL, and Lmax are discussed here. Performance Analyses of ATC-GHA for DRTS-SDF: The first ninety problem instances (small-scale data) are solved by both mathematical model as well as seven different ATC-GHA for DRTS-SDF in order to obtain the optimal solution and feasible solution, respectively, for the objective: TT, TWT, NT, WNT, OTD rate, 264 M. Vimala Rani and M. Mathirajan Fig. 1 Performance of seven different ATC-GHA for DRTS-SDF with respect to optimal: TT, TWT, NT, WNT, OTD rate, TEL, TWEL, and Lmax: over 90 problem instances TEL, TWEL, and Lmax. Using the solutions obtained on small scale data, this study computes RPD for each objective as per the Eq. (44) by considering the optimal solution as a benchmark solution. Then, for each of the scheduling objectives and each of the different ATC-GHA, this study calculates the average value of ninety small-scale problems’ RPD values. ARPD score of seven different ATC-GHA for each of the objectives are presented in Fig. 1. From this figure, it is observed based on solving small-scale problems that, by and large, ATC-GHA5, and ATC-GHA6 are performing better for the majority of the objectives considered here. The reason for the outperforming ATC-GHA may be due to the following: (a) The ATC rule (given by Li et al. (2010)) considered in ATC-GHA5 helps to compute exact slack time. This helps to avoid getting high priority for low Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 265 priority jobs. In addition, the ATC rule (given by Vimala Rani and Mathirajan (2016a)) considered in ATC-GHA6 ensures that if the slack is negative or the job is tardy, the ATC rule reduces not only to WSPT rule but also to weighted shortest tardy. (b) The BATC index rule (given by Mönch et al. (2006)) considered in ATC-GHA5 and ATC-GHA6 ensures that unavailable jobs are assigned low priority. Furthermore, this study evaluates seven different ATC-GHA w.r.t. estimated optimal solution (EOS) as (i) the mathematical models are extremely time-consuming, which is not suitable for large-scale data, and (ii) the findings observed based on small-scale data may not be generalized to say that this is true for large-scale data also. Accordingly, EOS is computed as per Rardin and Uzsoy (2001) for every 450 large scale data, and for each of the objective functions. By considering these estimated optimal solutions as a benchmark solution, RPD is computed as per the Eq. (44) for each problem instance and for each objective. Then, for every 45 problem configurations, ARPD is calculated using the corresponding RPD score obtained over ten problem instances per configuration, as per Eq. (45) for the objectives: TT, TWT, NT, WNT, OTD rate, TEL, TWEL & Lmax. Due to the brevity of the study, these results are not presented. However, the result obtained for the objective TT is given in Table 6, as an example. However, the average of the RPD values over entire problem instances is computed as per Eq. (45) by considering ‘N’ as 450 for each objective and is presented in Fig. 2. It is observed from Table 6 and Fig. 2 that, for large-scale data, by and large, ATCGHA5 and ATC-GHA6 are outperforming ATC-GHA for most of the scheduling objectives. It endorses the observation obtained from small-scale data. This study triangulates the findings observed from the ARPD analysis by using another performance measure, IRANK, over 450 problem instances. Accordingly, a [450×7] ranking matrix, indicating a rank for each of the seven different ATCGHA, is constructed for every objective considered here. Using the ranking matrix [450×7], a frequency ranking matrix [7×7] indicating how many times each of the proposed ATC-GHA performed to have a particular rank is computed for every objective. A sample frequency ranking matrix for the scheduling objective (minimizing TT) is given in Table 7. Using the computed frequency ranking matrix [7×7] of each of the scheduling objectives, the IRANK of seven different ATCGHA is computed using the Eq. (46). The computed IRANK of seven different ATC-GHA w.r.t. to every objective is presented in Table 7. This Table indicates that ATC-GHA5 and ATC-GHA6 are outperforming ATC-GHA for most of the objectives. This is clearly endorsing the observation obtained from the analysis of ARPD score wr.t. large-scale data. Performance Analyses of ATC-GHA for DRTS-NPDF-MER Due to computational intractability in scheduling DRTS-NPDF-MER for real-life sized problems and the highly questionable nature of generalizability of better performing variants observed based on small-scale problem, the performance analyses are carried out only on large-scale data considering both performance measures: ARPD and IRANK. Similar to single diffusion furnace, for every objective, problem instance wise EOS Problem configuration J1.A1,D1 J1.A1,D2 J1,A1,D3 J1,A2,D1 J1,A2,D2 J1,A2,D3 J1,A3,D1 J1,A3,D2 J1,A3,D3 J2.A1,D1 J2.A1,D2 J2,A1,D3 J2,A2,D1 J2,A2,D2 J2,A2,D3 J2,A3,D1 J2,A3,D2 J2,A3,D3 J3,A1,D1 J3,A1,D2 J3,A1,D3 J3,A2,D1 J3,A2,D2 S.No 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 ARPD score of seven different ATC-GHA in comparison with EOS ATC-GHA1 ATC-GHA2 ATC-GHA3 ATC-GHA4 274.05 32.39 18.05 −0.74 861.21 103.36 126.04 127.91 1248.84 5.80 84.06 84.06 275.68 228.64 228.64 228.64 149.78 0.00 0.00 44.00 46.44 22.59 −3.41 71.92 765.59 76.29 42.73 60.98 47.90 0.45 −6.22 0.45 −13.33 0.00 0.00 0.00 105.90 7.65 4.77 4.77 261.05 44.00 11.12 8.12 1693.88 −2.47 −1.62 −0.76 31.12 13.48 0.23 8.20 972.17 873.77 −5.74 859.24 404.08 0.92 −1.67 48.71 82.26 5.48 −0.39 1.37 64.84 53.17 3.42 39.54 416.12 118.91 375.32 142.30 102.05 55.73 44.09 58.18 294.71 46.08 19.23 54.48 142.79 29.44 38.85 55.72 56.80 10.11 4.31 7.47 110.47 49.76 22.76 39.79 Table 6 Performance of seven different ATC-GHA for DRTS-SDF w.r.t. EOS ATC-GHA5 −0.25 126.04 84.06 228.64 0.00 −3.41 42.73 −6.22 0.00 4.77 11.12 −1.62 0.23 −5.74 −1.67 −0.39 3.42 375.32 44.41 18.32 38.85 4.67 22.76 ATC-GHA6 −0.25 126.04 84.06 228.64 0.00 −3.41 42.73 −6.22 0.00 4.77 11.12 −1.62 0.23 −5.74 −1.67 −0.39 3.42 375.32 44.41 20.59 40.50 4.67 22.76 (continued) ATC-GHA7 −0.25 −2.62 104.06 426.86 906.67 10.86 502.26 35.16 0.00 33.36 8.00 57.57 11.80 866.85 41.57 3.96 50.75 142.30 61.37 73.82 58.73 17.92 42.07 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 J3,A2,D3 J3,A3,D1 J3,A3,D2 J3,A3,D3 J4.A1,D1 J4.A1,D2 J4,A1,D3 J4,A2,D1 J4,A2,D2 J4,A2,D3 J4,A3,D1 J4,A3,D2 J4,A3,D3 J5,A1,D1 J5,A1,D2 J5,A1,D3 J5,A2,D1 J5,A2,D2 J5,A2,D3 J5,A3,D1 J5,A3,D2 J5,A3,D3 Table 6 (continued) 351.55 99.88 169.85 172.80 59.78 50.37 85.67 43.18 36.14 58.09 38.08 55.54 94.98 8.88 23.71 24.09 37.56 35.70 23.96 38.40 35.95 46.70 141.99 30.27 78.57 51.21 24.94 18.29 39.23 27.19 8.22 34.34 20.26 40.92 23.03 7.94 25.64 23.07 23.61 27.30 24.24 22.13 16.40 34.61 79.43 31.57 65.15 33.41 31.10 33.48 59.25 25.02 5.40 30.63 13.51 32.75 17.71 −3.27 36.26 13.33 32.34 24.87 18.22 16.40 11.60 10.51 144.26 21.68 76.41 63.67 29.32 32.61 46.41 27.86 7.74 30.89 23.77 38.17 16.78 4.09 40.59 16.32 38.72 27.36 17.62 19.31 18.21 34.19 77.62 31.57 67.30 33.41 32.15 30.61 58.34 27.42 9.00 35.04 12.43 32.90 18.01 −2.61 32.25 19.80 33.20 29.07 18.94 17.31 16.73 14.42 77.94 32.08 67.54 33.41 32.17 31.40 61.46 27.35 8.46 37.17 12.79 33.59 22.71 −2.91 29.74 28.19 26.31 29.92 23.36 20.79 20.91 15.95 71.82 31.87 67.19 60.41 47.45 35.17 72.62 40.22 13.46 46.05 21.44 34.55 48.84 9.12 45.23 28.03 41.10 40.76 31.69 17.18 26.71 21.05 268 M. Vimala Rani and M. Mathirajan Fig. 2 Performance of seven different ATC-GHA for DRTS-SDF w.r.t. estimated optimal: TT, TWT, NT, WNT, OTD rate, TEL, TWEL, and Lmax: over 450 problem instances is computed using seven objective values, and then using these EOS as benchmark solution, RPD is computed as per Eq. (44). Then, objective-wise, the score on ARPD for (i) 10 problem instances over each of the problem configuration, and (ii) 450 problem instances are computed for seven different ATC-GHA for DRTS-NPDFMER. To keep this chapter brief, the results developed from the solution obtained, similar to Table 6 and Fig. 2, are not presented here. Similar to single diffusion furnace, IRANK performance score, while scheduling DRTS-NPDF-MER over 450 problem instances, was computed to triangulate the analysis with ARPD score. To keep this chapter brief, the computed IRANK of seven different ATG-GHA for DRTS-NPDF-MER is not presented here. ATC-GHA1 ATC-GHA2 ATC-GHA3 ATC-GHA4 ATC-GHA5 ATC-GHA6 ATC-GHA7 Proposed ATC-GHAs Number of times the proposed ATC-GHAs for DRTS-SDF yielded relatively best TT with the rank position of 1 2 3 4 5 6 7 77 79 80 63 64 56 31 153 102 65 51 33 30 16 203 92 68 47 22 14 4 135 115 81 42 35 25 17 174 110 64 51 34 15 2 178 98 58 49 42 16 9 145 87 72 41 42 32 31 The ranking position, based on IRANK (over 450 problem instances), of seven different ATC-GHA for DRTS-SDF, the scheduling objective TT TWT NT WNT OTD rate TEL WTEL Lmax 7 7 7 7 7 6 7 3 4 3 3 3 3 4 3 4 1 4 4 4 4 1 5 1 5 5 5 5 5 5 4 4 2 1 2 2 2 2 1 2 3 2 1 1 1 3 2 5 6 6 6 6 6 7 6 6 Table 7 Performance of seven different ATC-GHA for DRTS-SDF, over 450 instances, w.r.t. IRANK 270 M. Vimala Rani and M. Mathirajan However, from the analyses of the ARPD scores with respect to configurationwise (10 problem instances per configuration) and irrespective of the configuration (i.e. overall 450 instances), it is observed that by and large, ATC-GHA5 and ATCGHA6 are continuously performing better for the majority of the objectives in the large-scale data. This observation was further endorsed by analyzing the IRANK score. The better performing ATC-GHA, observed based on the empirical analyses for both cases of DRTS-SDF and DRTS-NPDF-MER, are further confirmed by statistical analysis, and the same is discussed in the next section. 6.2 Statistical Analyses on the Performance of ATC-GHA for DRTS of Diffusion Furnaces In addition to empirical analyses presented here, this study conducts statistical analysis using SPSS software. Accordingly, this study computes descriptive statistics: mean, median, standard deviation, and 95% confidence interval, for seven different ATC-GHA for DRTS of SDF w.r.t. every objective considered here, using 450 objective values obtained corresponding to 450 problem instances and is given in Table 8. It is observed from Table 8 that ATC-GHA5 is continuously outperforming for all the objectives. Moreover, ATC-GHA6 is also performing better for most of the objectives. Further, for each objective, this study desires to check whether the distribution of RPD across seven different ATC-GHA for DRTS-SDF is the same or not. For that, this study conducts Kruskal-Wallis non-parametric test, and the results are presented in Table 9a, b. From Table 9a, it is observed that a statistically significant difference is there in the distribution of RPD values across seven different ATC-GHA, as the PValues is 0.000, which is less than 0.05. In addition, Table 9b clearly says that ATCGHA5 and ATC-GHA6 are relatively outperforming heuristic algorithms among seven different ATC-GHA for DRTS-SDF to optimize various objectives considered here. Similar to statistical analyses carried out on seven different ATC-GHA for DRTS of SDF were performed on seven different ATC-GHA for DRTS-NPDF-MER. Due to the brevity of the report, the statistical results obtained are not presented here. However, the inference obtained from the statistical results endorses the same inferences obtained with respect to seven different ATC-GHA for DRTS-SDF to optimize various objectives. From both the analyses, this study detected that by and large ATC-GHA5, and ATC-GHA6 are continuously performing better for both DRTS-SDF and DRTS-NPDF-MER to optimize most of the objectives considered in this study. This could be due to the job-priority-index and batch-priority-index computed by these ATC-GHA that could yield an exact value of the slack. And this helps to avoid getting high priority for low priority jobs or unavailable jobs. Further, for the tardy jobs, these rules give a high priority to the job, which has the largest weighted tardy. TT ATC-GHA1 ATC-GHA2 ATC-GHA3 ATC-GHA4 ATC-GHA5 ATC-GHA6 ATC-GHA7 ATC-GHA1 ATC-GHA2 ATC-GHA3 ATC-GHA4 ATC-GHA5 ATC-GHA6 ATC-GHA7 ATC-GHA1 ATC-GHA2 ATC-GHA3 ATC-GHA4 ATC-GHA5 ATC-GHA6 ATC-GHA7 NT TWT Scheduling Objective Proposed ATC-GHA Descriptive Statistics Mean Median 1392.99 407.00 1248.85 248.50 1209.36 239.00 1263.90 268.00 1225.41 239.00 1240.50 239.00 1322.84 278.50 7617.35 2169.00 5245.97 1235.50 5314.14 1169.00 5422.12 1393.50 5195.25 1137.00 5253.20 1137.00 5462.17 1405.00 30.61 17.50 25.98 13.00 25.99 13.00 26.42 13.00 25.80 13.00 25.46 13.00 26.46 13.00 Table 8 Descriptive statistics of the proposed seven different ATC-GHA for DRTS-SDF Standard Deviation 2005.15 1879.97 1840.64 1919.33 1862.93 1876.87 2002.29 11048.96 7769.11 7996.89 8125.54 7834.27 7889.45 8115.24 32.44 29.17 29.30 29.51 28.97 28.58 29.42 (continued) 95% Confidence Interval (1207.73,1578.25) (1075.15,1422.54) (1039.29,1379.42) (1086.57,1441.24) (1053.28,1397.53) (1067.09,1413.91) (1137.84,1507.84) (6596.58638.2) (4528.16,5963.79) (4575.28,6053) (4671.37,6172.87) (4471.41,5919.09) (4524.26,5982.13) (4712.38,6211.97) (27.61,33.61) (23.28,28.67) (23.28,28.69) (23.69,29.14) (23.12,28.48) (22.82,28.1) (23.75,29.18) WNT ATC-GHA1 ATC-GHA2 ATC-GHA3 ATC-GHA4 ATC-GHA5 ATC-GHA6 ATC-GHA7 ATC-GHA1 ATC-GHA2 ATC-GHA3 ATC-GHA4 ATC-GHA5 ATC-GHA6 ATC-GHA7 ATC-GHA1 ATC-GHA2 ATC-GHA3 ATC-GHA4 ATC-GHA5 ATC-GHA6 ATC-GHA7 TEL OTD rate Scheduling Objective Proposed ATC-GHA Table 8 (continued) Descriptive Statistics Mean Median 168.45 97.00 137.35 68.00 138.69 70.00 140.84 70.00 137.12 70.00 135.48 70.00 141.47 74.50 0.34 0.26 0.46 0.40 0.47 0.40 0.46 0.40 0.47 0.40 0.47 0.40 0.45 0.40 1470.75 460.00 1399.99 452.00 1370.57 428.00 1413.53 440.00 1384.05 429.00 1402.19 431.50 1472.40 438.00 Standard Deviation 178.90 154.93 156.85 157.83 154.15 152.96 157.89 0.27 0.26 0.26 0.26 0.26 0.25 0.26 2010.13 1908.26 1869.04 1939.35 1889.40 1909.40 2027.01 (continued) 95% Confidence Interval (151.92,184.98) (123.03,151.66) (124.2153.18) (126.26,155.42) (122.88,151.37) (121.34,149.61) (126.88,156.06) (0.32,0.37) (0.44,0.49) (0.44,0.49) (0.43,0.48) (0.45,0.49) (0.45,0.5) (0.42,0.47) (1285.03,1656.48) (1223.68,1576.3) (1197.88,1543.25) (1234.35,1592.72) (1209.48,1558.62) (1225.78,1578.61) (1285.12,1659.69) ATC-GHA1 ATC-GHA2 ATC-GHA3 ATC-GHA4 ATC-GHA5 ATC-GHA6 ATC-GHA7 ATC-GHA1 ATC-GHA2 ATC-GHA3 ATC-GHA4 ATC-GHA5 ATC-GHA6 ATC-GHA7 Table 8 (continued) Lmax TWEL 8049.60 6119.09 6245.64 6283.41 6108.99 6181.25 6317.45 70.20 70.53 68.33 70.32 69.12 70.52 73.04 2613.00 2274.50 2338.50 2386.00 2270.50 2270.50 2319.50 55.00 50.00 49.00 50.00 49.00 49.00 55.00 11078.13 7951.30 8194.66 8260.75 8017.37 8090.68 8277.90 56.54 62.41 61.50 62.20 62.19 63.19 63.59 (7026.06,9073.15) (5384.44,6853.74) (5488.51,7002.78) (5520.17,7046.65) (5368.24,6849.74) (5433.72,6928.77) (5552.62,7082.27) (64.97,75.42) (64.76,76.29) (62.64,74.01) (64.58,76.07) (63.37,74.86) (64.69,76.36) (67.16,78.91) ATC-GHA1 ATC-GHA2 ATC-GHA3 ATC-GHA4 ATC-GHA5 ATC-GHA6 ATC-GHA7 Total 450 450 450 450 450 450 450 3150 N 0 0 0 6 OTD Rate 243.242 0 6 TEL 70.504 (b): Mean Rank for the seven different ATC-GHA w.r.t the Objective TT TWT NT WNT OTD Rate 2060.41 2204.68 2032.64 2042.05 2165.02 1590.35 1511.50 1534.63 1525.30 1512.24 1350.76 1409.86 1451.84 1455.17 1432.73 1575.73 1567.92 1551.39 1558.45 1531.01 1362.28 1345.14 1423.02 1421.52 1391.54 1392.81 1373.55 1410.03 1419.16 1390.06 1696.16 1615.85 1624.96 1606.86 1605.89 0 0 6 6 6 6 WNT 160.027 Proposed ATC-GHA KruskalWallis H Degrees of freedom Asymp. Sig. (a): Test Statistics for the Objective TT TWT NT 213.202 293.054 165.270 Table 9 Kruskal-Wallis test results on seven different ATC-GHA for DRTS-SDF TEL 1822.38 1561.60 1431.27 1562.18 1443.44 1488.08 1719.55 0 6 TWEL 2032.40 1504.32 1476.61 1546.32 1406.56 1451.08 1611.20 TWEL 147.988 0 6 Lmax 1768.72 1608.47 1420.86 1581.76 1432.56 1496.05 1720.09 Lmax 62.566 Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 275 7 Conclusion A new research problem on DRTS of DF(s), considering the job and/or resource related RTE along with different job-arrival time, machine eligibility restriction, incompatible-job families, and non-agreeable release time & due-date to optimize one at a time the objectives: TT, TWT, NT, WNT, OTD rate, TEL, TWEL, and Lmax is considered in this study. (0-1) MILP models for DS-SDF, as well as DS-NPDFMER, considering each of the objectives, are proposed. LINGO Set Code for all the (0-1) MILP models is also developed. All the proposed models are validated by solving ninety small-scale data. Since the mathematical models are not amenable to solve large-scale data, this study proposed seven different ATC-GHA for (i) DRTS-SDF, and (ii) DRTS-NPDFMER. The performance evaluation of seven different ATC-GHA for (i) DRTS-SDF, and (ii) DRTS-NPDF-MER is carried out in comparison with (a) the optimal value for 90 small-scale data, and (b) the estimated optimal value for 450 large scale data, which are generated based on the experimental design, using ARPD and IRANK measure. In addition to the empirical analyses, this study employs statistical analyses: descriptive statistics and Kruskal Wallis test. From both the analyses, this study concludes that, by and large, ATC-GHA5 and ATC-GHA6 are outperforming GHA for DRTS-SDF as well as DRTS-NPDF-MER to optimize the objectives: TT, TWT, NT, WNT, OTD rate, TEL, TWEL, and Lmax, one at a time. This could be because the job-priority-index and batch-priority-index computed by these ATC-GHA that could provide an accurate value of slack. This helps to avoid getting high priority for low priority jobs or unavailable jobs. Further, for the tardy jobs, these rules give a high priority to jobs which have large weighted tardy. The important managerial implication stems from identifying better performing ATC-GHA that would help the production manager to generate an efficient schedule even if any unexpected real time events occur. Further, it helps to deliver the product on or before its due date, which helps to improve the performance of the semiconductor manufacturing industry. The bench mark solution procedure: estimated optimal solution procedure has its own limitations. Moreover, the problem instances used in this study are generated not collected from SM industry. Considering (a) completion time-based scheduling objectives, and (b) upstream/downstream operation(s) while doing DRTS of diffusion furnace(s) without the limitation listed above would be the future research direction in this area. References Balasubramanian, H., Monch, L., Fowler, J., & Pfund, M. (2004, January 01). Genetic algorithm based scheduling of parallel batch machines with incompatible job families to minimize total weighted tardiness. Int J Prod Res, 42(8), 1621–1638. 276 M. Vimala Rani and M. Mathirajan Bar-Noy, A., Guha, S., Katz, Y., Naor, J. S., Schieber, B., & Shachnai, H. (2009, March 01). Throughput maximization of real-time scheduling with batching. ACM Transactions on Algorithms, 5(2), 1–17. Beldar, P., & Costa, A. (2018, January 01). Single machine batch processing problem with release dates to minimize total completion time. International Journal of Industrial Engineering Computations, 9(3), 331–348. Bilyk, A., Mönch, L., & Almeder, C. (2014, October 23). Scheduling jobs with ready times and precedence constraints on parallel batch machines using meta heuristics. Computers & Industrial Engineering, 78, 175–185. Chen, L., Lu, C., Hui, X., & Li, L. (2013, December 01). Learning-based adaptive dispatching method for batch processing machines. In Proceedings of the 2013 Winter Simulation Conference, pp. 3756–3765. Cheng, H. C., Chiang, T. C., & Fu, L. C. (2008, December 01). A memetic algorithm for parallel batch machine scheduling with incompatible job families and dynamic job arrivals. Proceedings of IEEE International Conference on Systems, Man and Cybernetics, 2008, 541– 546. Chiang, T.-C., Cheng, H.-C., & Fu, L.-C. (2008, November 01). An efficient heuristic for minimizing maximum lateness on parallel batch machines. Proceedings of the Eighth International Conference on Intelligent Systems Design and Applications, 2, 621–627. Chiang, T.-C., Cheng, H.-C., & Fu, L.-C. (2010, December 01). A memetic algorithm for minimizing total weighted tardiness on parallel batch machines with incompatible job families and dynamic job arrival. Computers & Operations Research, 37(12), 2257–2269. Dirk, R., & Monch, L. (2006, January 01). Multiobjective scheduling of jobs with incompatible families on parallel batch machines. Lecture Notes in Computer Science, 3906, 209–221. Fidelis, M. B., & Arroyo, J. E. C. (2017, October 01). Meta-heuristic algorithms for scheduling on parallel batch machines with unequal job ready times. Proceedings of the IEEE International Conference on Systems, Man and Cybernetics, 2017, 542–547. Guo, C., Jiang, Z., & Hu, H. (2010, December 01). A hybrid ant colony optimization method for scheduling batch processing machine in the semiconductor manufacturing. IEEE International Conference on Industrial Engineering & Engineering Management (IE&EM), 2010, 1698– 1701. Hildebrandt T, Heger J, Scholz-Reiter B (2010, July 07). Towards improved dispatching rules for complex shop floor scenarios—A genetic programming approach. In Proceeding of the GECCO’10, July 7–11, Portland, Oregon, USA. Retrieved April 3, 2020, from http:// www.genetic-programming.org/hc2010/6-Hildebrandt/Hildebrandt-Paper.pdf Jia, W., Jiang, Z., & Li, Y. (2013a, August 01). Closed loop control-based real-time dispatching heuristic on parallel batch machines with incompatible job families and dynamic arrivals. International Journal of Production Research, 51(15), 4570–4584. Jia, W., Jiang, Z., & Li, Y., (2013b, August 01). A job-family-oriented algorithm for re-entrant batch processing machine scheduling. In IEEE International Conference on Automation Science and Engineering, pp. 1022–1027. Jung, C., Pabst, D., Ham, M., Stehli, M., & Rothe, M. (2014, August 01). An effective problem decomposition method for scheduling of diffusion processes based on mixed integer linear programming. IEEE Transactions on Semiconductor Manufacturing, 27(3), 357–363. Kim, Y. D., Joo, B. J., & Choi, S. Y. (2010, January 01). Scheduling wafer lots on diffusion machines in a semiconductor wafer fabrication facility. IEEE Transactions on Semiconductor Manufacturing, 23(2), 246–254. Koo, P.-H., & Moon, D. H. (2013, January 01). A review on control strategies of batch processing machines in semiconductor manufacturing. IFAC Proceedings, 46(9), 1690–1695. Kurz, M.E. (2003, May 22). On the structure of optimal schedules for minimizing total weighted tardiness on parallel, batch-processing machines. In Proceedings of 10th IE Research Conference, Portland, pp. 1–5. Prescriptive Analytics for Dynamic Real Time Scheduling of Diffusion Furnaces 277 Kurz, M. E., & Mason, S. J. (2008, January 01). Minimizing total weighted tardiness on a batchprocessing machine with incompatible job families and job ready times. International Journal of Production Research, 46(1), 131–151. Li, L., &Qiao, F. (2008, August 01). ACO-based scheduling for a single batch processing machine in semiconductor manufacturing. In 4th IEEE International Conference on Automation Science and Engineering, pp. 85–90. Li, L., Qiao, F., & Wu, Q.D. (2009, August 01). ACO-based scheduling of parallel batch processing machines to minimize the total weighted tardiness. In 5th Annual IEEE Conference on Automation Science and Engineering, pp. 280–285. Li, L., Qiao, F., & Pan, G. (2010, August 17). ACO-based multi-objective scheduling of identical parallel batch processing machines in semiconductor manufacturing. INTECH Open Access Publisher. Retrieved April 3, 2020, from http://www.intechopen.com/books/futuremanufacturing-systems/aco-based-multi-objective-scheduling-of-identical-parallel-batchprocessing-machines-in-semiconducto Li, Y., Jiang, Z., & Jia, W. (2017, January 02). API-based two-dimensional dispatching decisionmaking approach for semiconductor wafer fabrication with operation due date-related objectives. International Journal of Production Research, 55(1), 79–95. Malve, S., & Uzsoy, R. (2007, January 01). A genetic algorithm for minimizing maximum lateness on parallel identical batch processing machines with dynamic job arrivals and incompatible job families. Computers and Operations Research, 34(10), 3016–3028. Mansoer, P., & Koo, P.-H. (2015, January 01). A batching strategy for batch processing machine with multiple product types. Journal of Industrial and Intelligent Information, 3(2), 138–142. Mathirajan, M., & Vimalarani, M. (2012, December 13). Scheduling a BPM with incompatible job-families and dynamic job-arrivals. In Proceedings of the 2012 IEEE International Conference on Industrial Engineering and Engineering Management, pp. 622–626. Retrieved April 3, 2020, from https://www.researchgate.net/publication/ 265597365_Scheduling_a_BPM_with_Incompatible_Job-Families_and_Dynamic_JobArrivals Mathirajan, M., & Sivakumar, A. I. (2006, July 01). A literature review, classification and simple meta-analysis on scheduling of batch processors in semiconductor. The International Journal of Advanced Manufacturing Technology, 29, 990–1001. Mathirajan, M., Sivakumar, A. I., & Kalaivani, P. (November, 2004). An efficient simulated annealing algorithm for scheduling burn-in oven with non-identical job-sizes. The International Journal of Applied Management and Technology, 2(2), 117–138. Mehta, S. V., & Uzsoy, R. (1998, February 01). Minimizing total tardiness on a batch processing machine with incompatible job families. IIE Transactions, 30(2), 165–178. Mönch, L., & Roob, S. (2018, January 01). A matheuristic framework for batch machine scheduling problems with incompatible job families and regular sum objective. Applied Soft Computing, 68, 835–846. Mönch, L., Balasubramanian, H., Fowler, J. W., & Pfund, M. E. (2005, November 01). Heuristic scheduling of jobs on parallel batch machines with incompatible job families and unequal ready times. Computers & Operations Research, 32(11), 2731–2750. Mönch, L., Zimmermann, J., & Otto, P. (2006, April 01). Machine learning techniques for scheduling jobs with incompatible families and unequal ready times on parallel batch machines. Engineering Applications of Artificial Intelligence, 19(3), 235–245. Monch, L., Fowler, J., Dauzère-Pérès, S., Mason, S., & Rose, O. (2011, January 07). A survey of problems, solution techniques, and future challenges in scheduling semiconductor manufacturing operations. Journal of Scheduling, 14(6), 583–599. Mönch, L., Fowler, J., & Mason, S. J. (2013). Production planning and control for semiconductor wafer fabrication facilities: Modeling, analysis, and systems. New York, NY: Springer. Mönch, L., Chien, C.-F., Dauzère-Pérès, S., Ehm, H., & Fowler, J. W. (2018, July 03). Modelling and analysis of semiconductor supply chains. International Journal of Production Research, 56(13), 4521–4523. 278 M. Vimala Rani and M. Mathirajan Park, H., & Banerjee, A. (2011, December 01). A new dynamic scheduling for batch processing systems using stochastic utility evaluation function. Proceedings of the 2011 Winter Simulation Conference, pp. 2302–2319. Pinedo, M. (2012). Scheduling: Theory, algorithms, and systems. New York: Springer. Pirovano, G., Ciccullo, F., Pero, M., & Rossi, T. (2020, January 01). Scheduling batches with time constraints in wafer fabrication. International Journal of Operational Research, 37(1), 1. Rardin, R. L., & Uzsoy, R. (2001, January 01). Experimental evaluation of Heuristic optimization algorithms: A tutorial. Journal of Heuristics, 7, 261–304. Rocholl, J., Monch, L., & Fowler, J. W. (2018, December 01). Electricity power cost-aware scheduling of jobs on parallel batch processing machines. In Proceedings - Winter Simulation Conference, pp. 3420–3431. Sarin, S. C., &Shikalgar, S. T. (2001, August 25). Reduction of average cycle time at a wafer fabrication facility. GaAs MANTECH, Inc, Grado Department of Industrial and Systems Engineering, Virginia Tech. Retrieved April 3, 2020, from https://pdfs.semanticscholar.org/ d3fd/2c0e192b6bcde5d2166ac7d91a2e81ac0917.pdf SGSR: Semiconductor Global Sales Report. (2020). Retrieved August 28, 2020, from https:// www.statista.com/statistics/266973/global-semiconductor-sales-since-1988/ Sun, C., & Rose, T. (2015, January 01). Supply Chain complexity in the semiconductor industry: Assessment from system view and the impact of changes. IFAC PapersOnline, 48(3), 1210– 1215. Uzsoy, R. (1995, October 01). Scheduling batch processing machines with incompatible job families. International Journal of Production Research, 33(10), 2685–2708. Vimala Rani, M., & Mathirajan, M. (2015, December 01). Dynamic scheduling of diffusion furnace in semiconductor manufacturing with real time events. IEEE International Conference on Industrial Engineering and Engineering Management. https://doi.org/10.1109/ IEEM.2015.7385617. Vimala Rani, M., & Mathirajan, M. (September 02, 2016a). Performance evaluation of ATC based greedy heuristic algorithms in scheduling diffusion furnace in wafer fabrication. Journal of Information and Optimization Sciences, 37(5), 717–762. Vimala Rani, M., & Mathirajan, M. (2016b, May 05). Dynamic scheduling of diffusion furnace in semiconductor manufacturing with resource related real time events. In Analytics in operations/supply chain management (pp. 381–396). New Delhi: I.K International Publishing House. Vimala Rani, M., & Mathirajan, M. (2016c, January 01). Multi objective dynamic real-time scheduling of batch processing machine. International Journal of Operations and Quantitative Management, 22(1), 53–73. Vimala Rani, M., & Mathirajan, M. (2020a, June 01). Performance evaluation of due-date based dispatching rules in dynamic scheduling of diffusion furnace. OPSEARCH, 57(2), 462–512. Vimala Rani, M., & Mathirajan, M. (2020b, January 01). Meta-heuristics for dynamic real time scheduling of diffusion furnace in semiconductor manufacturing industry. International Journal of Industrial and Systems Engineering, 34(3), 365–395. Wang, R., & Fei, S. (2014, January 01). Rescheduling: External environment-related real-time events. IFAC Proceedings Volumes, 47(3), 10743–10747.