Workshop on Supply Chain Analytics
− Honoring David Simchi-Levi on the
Occasion of His 60th Birthday
Sponsored by:
October 15-16, 2015
University of California, Berkeley
The Department of Industrial
Engineering and Operations
Research, University of California,
Berkeley
The MIT Forum for Supply Chain
Innovation, Massachusetts Institute
of Technology
Program Rundown
All academic talks take place in 290 Hearst Memorial Mining Building (HMMB)
Time
Activity
Day 1 (Thursday, October 15, 2015)
8:15 am – 9:15 am
Registration (HMMB Lobby)
9:15 am – 9:25 am
Welcoming Remarks by Philip Kaminsky
9:25 am – 10:00 am
Special Remarks by Pierre Haren, Wallace J. Hopp, and
Narendra P. Mulani
10:00 am – 10:45 am
Academic Talk by Opher Baron and Oded Berman
10:45 am – 11:15 am
Coffee break (HMMB Lobby)
11:15 am – 12:00 noon
Academic Talk by Xiuli Chao
12:00 noon – 1:30 pm
Lunch (HMMB Lobby)
1:30 pm – 2:15 pm
Academic Talk by Georgia Perakis
2:15 pm – 3:00 pm
Academic Talk by Mark S. Daskin
3:00 pm – 3:30 pm
Coffee break (HMMB Lobby)
3:30 pm – 4:15 pm
Academic Talk by Guillermo Gallego
4:15 pm – 5:00 pm
Academic Talk by Stephen C. Graves
5:00 pm
Cocktails and Banquet (Berkeley Faculty Club)
Day 2 (Friday, October 16, 2015)
9:15 am – 10:00 am
Academic Talk by Caro Felipe
10:00 am – 10:45 am
Academic Talk by Nir Halman
10:45 am – 11:15 am
Coffee break (HMMB Lobby)
11:15 am – 12:00 noon
Academic Talk by Zvi Drezner
12:00 noon – 1:30 pm
Lunch (HMMB Lobby)
1:30 pm – 2:15 pm
Academic Talk by Max Shen
2:15 pm – 3:00 pm
Academic Talk by Chung Piaw Teo
3:00 pm – 3:10 pm
Closing Remarks by Chung-Lun Li
1
Special Remarks
Day 1, 9:25 am – 10:00 am
Pierre Haren
Vice-President, Global Leader Advanced Analytics, GBS, IBM and former founder of ILOG
Wallace J. Hopp
Senior Associate Dean for Faculty & Research and Herrick Professor of Business, Stephen M.
Ross School of Business, University of Michigan
Narendra P. Mulani
Chief Analytics Officer, Accenture Analytics
2
Academic Talk by Opher Baron and Oded Berman
Day 1, 10:00 am – 10:45 am
Strategic Idling and Dynamic Scheduling in Open-Shop Service Network:
Case Study and Analysis
Opher Baron, Oded Berman, D. Krass (Rotman School of Management, University of Toronto),
and J. Wang (Nanyang Business School, Nanyang Technological University)
Abstract:
In Open-shop service networks customers would like to obtain service from a set of stations,
most of them without a specific order. This paper is motivated by XYZ (not the real name), a
company in the healthcare service industry that operates a stochastic open shop network where
the stations of the networks administer medical tests that customers can take within several hours
of the same day. According to senior management of XYZ there are two types of complaints
about the service that customers of XYZ experience. One is with respect to the total time that
customers spend in the system; the other is with respect to the long waiting time at a specific
station. In fact the company believes that customers get upset when they wait more than 20
minutes for a particular station (such customers appear on the computer screen of the schedulers
with red faces).
We focus in this paper on two types of service levels: the more traditional macro-level
measures such as minimizing total waiting time or total system time (waiting plus service times)
or minimizing total tardiness, and the “micro-level” measure of reducing excessive long waits at
any individual workstation within the process. The only paper we are aware of that discusses
systematically and analytically micro service level is [1] where a strategic idling (SI) scheduling
policy is suggested.
The idea behind SI is that when a downstream station is very congested operating the
upstream station in a normal rate may increase the congestion at the downstream station. Instead,
idling the upstream station until the downstream station is less congested could be beneficial.
Therefore while work-conserving policies are optimal for macro-level measures, scheduling
policies with SI might be helpful for the micro-level measure. In [1] we showed the benefits of
SI for the two stations in tandem network where customers arrive to the network according to a
Poisson process and services at the stations are exponentially distributed. In the current paper we
examine whether similar ideas can be applied to a much more complex environment of a
stochastic open shop network.
There was no official policy of using SI in XYZ. However using the empirical data we found
statistical evidence that SI is in fact used by the schedulers to effectively manage the micro-level
measure. This SI was done using only intuition of the schedulers of XYZ. We provide in this
paper an efficient way to combine the SI and Dynamic Scheduling Policies (DSPs)so that the
resulting policies can simultaneously address both macro- and micro-level measures.
For deciding which customer should be assigned to the next freed-up station we use 6 DSPs
that include among others rules such as: “Longest System time first” and “Longest Current
Waiting time first”. In all of the 6 DSPs used the station that is just freed-up and has the highest
remaining workload is assigned to a waiting customer.
3
Since the stochastic open-shop networks are very difficult to analyze analytically, we
developed two simulation models. The first simulation model is based on the empirical data (ED)
for arrivals and service times. We compared the micro and macro service levels for the following
policies: ED; ED with no idling; the six DSPs with no idling and the six DSPs with SI. The main
findings are: (1) ED with no idling results in better macro service levels than ED but with much
worse micro-level performance, (2) The DSPs with no idling are much better than ED in their
macro service levels but perform worse on micro-level than ED, (3) The DSPs with SI result in
worse (but not by a lot) macro service levels than those without idling but are much better than
the DSPs without SI and quite close to the ED policy in the micro-level performance. The second
simulation model is based on randomly generated open-shop networks aiming to show benefits
of using SI with DSPs for general networks. The results obtained with the second simulation are
in line with those of the first model and show that combining DSPs with SI is a promising
strategy in general stochastic open-shop environments.
Reference
[1] O. Baron, O. Berman, D. Krass and J. Wang. Using strategic idleness to improve customer
service experience in service networks. Operations Research, 2014, 62(1), 123-140.
4
Academic Talk by Xiuli Chao
Day 1, 11:15 am – 12:00 noon
Nonparametric Algorithm for Joint Pricing and Inventory Control with LostSales and Censored Demand
Xiuli Chao (Department of Industrial and Operations Engineering, University of Michigan); joint
work with Beryl Chen and Cong Shi
Abstract:
We consider the classic joint pricing and inventory control problem with lost-sales and censored
demand in which the customer response to selling price and the demand distribution is unknown
to the firm a priori. The major difficulty in this problem lies in that, the estimate of the expected
profit function from data can never be unimodal, even though the expected profit function is
unimodal under certain conditions. We develop a data-driven algorithm that conducts active
exploration and exploitation in carefully designed cycles, and show it converges to the optimal
policy when the planning horizon becomes long. The regret rate of the algorithm is also given.
5
Academic Talk by Georgia Perakis
Day 1, 1:30 pm – 2:15 pm
The Impact of Analytics on Promotion Pricing
Georgia Perakis (Sloan School of Management, Massachusetts Institute of Technology); joint
work with Lennart Baardman (ORC PhD student), Maxime Cohen, (recently graduated ORC
PhD student), Swati Gupta (ORC PhD student), Jeremy Kalas (EECS Undergraduate), Zachary
Leung (recently graduated ORC PhD student), Danny Segev (Visiting Scholar ORC/MIT Sloan
from U. Haifa), Kiran Panchamgam (Oracle RGBU), and Anthony Smith (formerly from Oracle
RGBU)
Abstract:
Pricing has been a field that has seen a lot of exciting developments in the recent years. A
particular area of pricing that has recently emerged is promotion pricing. In many important
settings, promotions are a key instrument for driving sales and profits. Important examples
include promotions in grocery retail among others. The Promotion Optimization Problem (POP)
is a challenging problem as the retailer needs to decide which products to promote, what is the
depth of price discounts, when to schedule the promotions and how to promote the product.
In this talk we will discuss how analytics can have a key impact on promotion pricing. This
presentation will reflect our ongoing collaboration over the past few years with Oracle RGBU.
We will describe the journey we took with them on introducing analytics tools for promotion
planning in the grocery industry. We will describe optimization models we have built in order to
determine which products to promote and when, as well as which vehicles to use to promote
each product (e.g. flyer versus radio announcements among others) and how deeply to promote
each product.
An important consumer behavior we will incorporate and which is a direct consequence of
promotions in grocery retail is that consumers stockpile the products on promotion and then
experience promotion fatigue after the promotion ends. Therefore, as a first step, we study
general classes of demand functions that capture this effect and can be directly estimated from
data. Using these demand functions, we model and study the promotion planning problem
through an optimization formulation. Unfortunately, the underlying formulation even for a single
product is NP-hard and highly nonlinear. We will first propose a linear approximation and show
how to solve the problem efficiently as a linear programming (LP) problem. We will illustrate
how this approximation idea has applications in many areas beyond pricing. We will discuss
analytical bounds on the accuracy of this LP approximation relative to exact problem solution.
We will also consider a graphical representation of the problem which will allow us to employ a
Dynamic Programming (DP) solution approach as an alternative. We will discuss the tradeoffs
between the two approaches (LP vs DP). Furthermore, we will incorporate in the optimization
formulations, apart from the pricing aspect, how to decide which vehicle to use each time in
order to promote which product. This further complicates the problem. We will introduce greedy
and integer optimization ideas in order to solve the vehicle selection problem in a tractable way.
These methods are computationally efficient and hence are easy to use in practice. We will also
discuss some performance guarantees for these methods.
6
Together with our industry collaborators from Oracle Retail, our framework allows us to
develop a tool which can help supermarket managers to better understand promotions by testing
various strategies and business constraints. We show that the formulation we propose solves fast
in practice using actual data from a grocery retailer and that the accuracy is high. We calibrate
our models using actual data and determine that they can improve profits by 3% just by
optimizing the promotion schedule and up to 5% by slightly modifying some business
requirements.
7
Academic Talk by Mark S. Daskin
Day 1, 2:15 pm – 3:00 pm
Mitigating Capacity Constraints in Facility Location Models
Mark S. Daskin and Kayse Lee Maass (Industrial and Operations Engineering, University of
Michigan)
Abstract:
Hard capacity constraints have been used for decades in facility location modeling and planning.
In this talk, we will argue that such constraints are unrealistic as a variety of operational tools can
be used to extend capacity in the short term. We will focus on the use of inventory as a short
term mitigation strategy as other strategies – overtime and outsourcing, for example – are easier
to model. At the same time, massive amounts of data are now available for planners. The models
we present enable the use of such databases.
We will present a basic model in which demands from a node must be assigned to a single
facility for all time intervals under consideration by the model. We show that the location
decisions made without allowing inventory to mitigate the capacity constraints are different, and
in some cases significantly worse, than those suggested by models that allow inventory
accumulation to mitigate capacity constraints. We then extend this model to allow for day of the
week assignments which can account for time-varying demands or for facilities that are only
open on selected days of the week.
8
Academic Talk by Guillermo Gallego
Day 1, 3:30 pm – 4:15 pm
Do Consumers Benefit from Dynamic Pricing?
Guillermo Gallego (Department of Industrial Engineering and Operations Research, Columbia
University) and Ningyuan Chen (Yale University)
Abstract:
We have asked this question to seasoned practitioners and leading academics and the answer is
often "I don't know, but my intuition is that if the seller is benefiting from dynamic pricing then
it may be because more surplus is being extracted from consumers." This perception may even
prevent the use of dynamic pricing if companies feel that it may hurt their image in the eyes of
consumers. It was precisely the expression of this sentiment that leads us to this investigation.
We asked ourselves: Do consumers prefer random prices? In other words, is the consumer
surplus a decreasing convex function of price? Is the price an increasing concave function of cost?
Is the composite function a decreasing convex function of cost? And if this is true, is dynamic
pricing (which responds to random changes in marginal cost of capacity), better than fixed
pricing for consumers? We show that the answers are mostly yes, so dynamic pricing is win-win
most of the time, but there are some exceptions as we will describe in our talk.
9
Academic Talk by Stephen C. Graves
Day 1, 4:15 pm – 5:00 pm
Modeling and Design of a Material Recovery Facility
Stephen C. Graves (Sloan School of Management, Massachusetts Institute of Technology); joint
work with Timothy Gutowski (MIT), Anne Raymond (MIT), and Mariapaola Testa (MIT)
Abstract:
The production of municipal solid waste (MSW) has increased world-wide, in both developed
and developing countries. In developed countries, MSW is usually sent to material recovery
facilities (MRFs), which use mechanical and manual sorting units to extract valuable
components. In this ongoing work, we formulate and develop a network flow model to represent
an MRF that sorts wastes with various sortation processes and with recirculating streams. For
each material flow in the system, we define a matrix to describe the sorting process. We use data
from an operating MRF to estimate the sortation parameters for the model, as well as to
characterize the input waste stream. We then formulate a genetic algorithm (GA) that generates
alternative configurations of an MRF for a given set of sorting units with known separation
parameters and selects those with highest profit and efficiency. The GA incorporates a heuristic
for personnel allocation to manual units. We report on our experience to date experimenting with
the model.
10
Academic Talk by Caro Felipe
Day 2, 9:15 am – 10:00 am
Managing Online Content to Build a Follower Base
Caro Felipe (Anderson School of Management, UCLA); joint work with Victor Martínez-deAlbéniz (IESE Business School)
Abstract:
Content providers typically manage a dual objective of generating interest for current followers
and at the same time reaching out to new audiences that may become repeat visitors. The pace at
which content is created must thus take into account how much it contributes to maintain the
follower base. We formulate a simple model to study follower base build-up dynamics under the
assumption that the attractiveness of past content decays over time. Using stochastic dynamic
programming we develop heuristics for content release and an upper bound to assess
performance. We then apply our model to the case of blogs: we analyze four blogs and their
daily traffic for over 3 years. Our findings show that adjusting posting frequency to the optimum
can raise total traffic levels by generating a larger follower base.
11
Academic Talk by Nir Halman
Day 2, 10:00 am – 10:45 am
Multi-Parameter Approximation Schemes for Hard Newsvendor Problems
Nir Halman (Jerusalem School of Business Administration, The Hebrew University of Jerusalem)
Abstract:
For every given real value
, a Fully Polynomial Time Approximation Scheme (FPTAS)
computes in polynomial time (in both the input size and
) a feasible solution that is close to
the optimal solution within ratio . As can be chosen arbitrarily small, and the running time is
polynomial, FPTASs are considered as the “Holy grail” of approximation algorithms.
In this talk I will study 3 hard variants of the classical newsvendor problem, each of which
cannot admit an FPTAS. However, by expanding the notion of FPTASs to multi-parameter
approximation schemes, it is possible to compute for them provably near-optimal feasible
solutions.
The talk is based upon 3 papers by (a subset of) Nir Halman, Giacomo Nannicini, Jim Orlin
and David Simchi-Levi.
12
Academic Talk by Zvi Drezner
Day 2, 11:15 am – 12:00 noon
An Iterative Procedure for Solving Non-Linear Programs
Zvi Drezner and Pawel Kalczynski (Steven G. Mihaylo College of Business and Economics,
California State University-Fullerton)
Abstract:
Non-linear programming problems of minimizing a convex objective function subject to convex
constraints are convex, have one local optimum which is the global one, and can be solved by
numerous approaches and canned programs. Non-convex programs such as a maximization of a
convex objective subject to constraints which are outside or on the outsides of convex regions
usually have many local optima and are generally difficult to solve. Many such problems are
solved by global optimization methods which exploit the special structure of the problem and
usually require excessive processing times that restrict the size of the problems that can be
practically solved. We found that such problems can be solved iteratively by a multi-start
heuristic approach based on solving a sequence of linear programs. It is interesting that this
general approach is suitable for solving non-convex problems but cannot be applied for solving
convex problems. Our iterative approach is much faster than a direct multi-start approach (one to
three orders of magnitude) and provided better results on four test problems and 116 instances.
Since the iterative procedure is much faster and provides better results, the multi-start approach
can be implemented with more starts providing excellent results in relatively short computing
times.
13
Academic Talk by Max Shen
Day 2, 1:30 pm – 2:15 pm
Service Region Design for Urban Electric Vehicle Sharing Systems
Max Shen (Department of Industrial Engineering and Operations Research, and Department of
Civil and Environmental Engineering, University of California-Berkeley)
Abstract:
Emerging collaborative consumption business models have shown promise in both generating
business opportunities and enhancing efficient use of resources. In the transportation domain, car
sharing models are adopted at mass scale in major metropolitan areas worldwide. This mode of
servicized mobility bridges the resource efficiency of public transit and the flexibility of personal
transportation. Beyond significant potential to reduce car ownership, car sharing shows promise
in supporting adoption of fuel efficient vehicles, such as electric vehicles (EVs), due to these
vehicles special cost structure with high purchase but low operating costs. Recently, it has
become a trend for key players in the car sharing business, such as Car2Go and Autolib, to
employ EVs in an operations model that accommodates one-way trips. On the one hand, the oneway model brings about significant improvements in coverage of travel needs and therefore
adoption potential, compared with the conventional round-trip-only model (advocated by ZipCar,
for example). On the other hand, it poses tremendous planning and operational challenges. In this
work, we study the planning problem faced by service providers in designing the geographical
service region in which to operate the service. This decision encompasses the trade-off between
maximizing customer catchment by covering travel needs, and controlling fleet operations costs.
We develop a mathematical programming model that incorporates details of both customer
adoption behavior and fleet management (including EV repositioning and charging) operations
under spatially-imbalanced and time-varying travel patterns. To address inherent planning
uncertainty with regard to adoption patterns, we employ a distributionally-robust optimization
framework that informs robust decisions to avoid possible ambiguity (or lack) of data.
Mathematically, the problem can be approximated by a mixed integer second-order cone
program, which is computationally- tractable with practical scale data. Applying this approach to
the case of Car2Go’s service in San Diego, California, with real operations data, we address a
number of planning questions and suggest potential for future development of the service.
14
Academic Talk by Chung Piaw Teo
Day 2, 2:15 pm – 3:00 pm
Risk Exposure Index Revisited − Prioritising and Mitigating Disruption Risks
in Supply Chains
Chung Piaw Teo (Department of Decision Sciences, National University of Singapore); joint
work with Zhenzhen Yan and Yini Sarah Gao (National University of Singapore)
Abstract:
In recent years, the concept of REI introduced by David has slowly gained traction among supply
chain professionals to study risk management. In this paper, we propose an extension to the
concept and show that the new index can model the supply chain time to survive distribution
profile more accurately. Furthermore, we propose a new model to design mitigation strategy
using the related time to recovery concept.
15
Organizing Committee
Philip Kaminsky (Chair), University of California, Berkeley
Chung-Lun Li (Co-chair), The Hong Kong Polytechnic University
Xin Chen, University of Illinois at Urbana-Champaign
Youhua (Frank) Chen, City University of Hong Kong
Ye Lu, City University of Hong Kong
Max Shen, University of California, Berkeley
Miao Song, The Hong Kong Polytechnic University
16
Venue
Registration and Academic Talks: 290 Hearst Memorial Mining Building
Banquet: The Faculty Club
Hearst Memorial Mining Building
The Faculty Club