UNIVERSITEIT GENT FACULTEIT ECONOMIE EN BEDRIJFSKUNDE ACADEMIEJAAR 2011 – 2012 LINKING THE VEHICLE ROUTING PROBLEM AND THE CREW SCHEDULING PROBLEM: PAST, PRESENT AND FUTURE RESEARCH Masterproef voorgedragen tot het bekomen van de graad van Master of Science in de Bedrijfseconomie Jordy Batselier onder leiding van Prof. dr. Broos Maenhout PERMISSION Ondergetekende verklaart dat de inhoud van deze masterproef mag geraadpleegd en/of gereproduceerd worden, mits bronvermelding. Undersigned declares that the contents of this thesis may be consulted and/or reproduced, provided acknowledgment. Jordy Batselier II Acknowledgements The thesis in front of you is written on behalf of Ghent University, the university at which I am proud to be a student for almost 6 years now. In my first 5 years I successfully waded through the Engineering program. Although specialized in the field of construction, I also chose to broaden my purely technical horizon and signed up for a minor in Business Administration. It was there that I first encountered and learned to greatly appreciate the area of Operations Research. Also my interest in economics was strongly sparked, which was the impetus to begin a complementary study in Business Economics. When looking for an interesting subject for my new master’s thesis, I was immediately attracted by the topics presented by the department of Operations Management. The eventual subject of this thesis was clearly my favorite given my previously developed interest in transportation, and I therefore want to thank my promotor prof. dr. Broos Maenhout very much for willing to adapt the original subject for me (since it was actually intended for a two-year Business Engineering thesis) and give me the necessary guidance. The writing of this thesis has also further contributed to my affection for Operations Research and thus played a significant role for the endeavor of commencing a PhD at the department of Operations Management. I would also like to thank my parents and my girlfriend’s parents for ensuring the necessary peace and quiet while writing my thesis during weekends, my friends for asking me about my thesis so that the discussions that followed could lead to new insights on my part, and last but certainly not least my girlfriend Lynn for taking care of me, supporting me and (sometimes) putting up with me during the busy days of writing this thesis. Jordy Batselier Ghent, May 2012 III Contents Acknowledgements ................................................................................................................... III Abbreviations ............................................................................................................................ VI Definitions ................................................................................................................................ VII List of figures ........................................................................................................................... VIII List of tables ............................................................................................................................ VIII 0. Introduction ......................................................................................................................... 1 1. Past ..................................................................................................................................... 5 1.1. Individual routing and scheduling problems .................................................................. 5 1.1.1. The Vehicle Routing Problem (VRP) ..................................................................... 6 1.1.2. The Vehicle Scheduling Problem (VSP) ................................................................ 8 1.1.2.1. Single depot case........................................................................................... 8 1.1.2.2. Multiple depot case ........................................................................................ 9 1.1.3. The Crew Scheduling Problem (CSP) ................................................................... 9 1.2. The planning process in transportation companies ..................................................... 11 1.3. The scheduling problem ............................................................................................. 16 1.3.1. Traditional approach ........................................................................................... 17 1.3.2. Defining the Vehicle and Crew Scheduling Problem (VCSP) ............................... 17 1.3.2.1. Public transport context ................................................................................ 18 1.3.2.2. Distribution context ....................................................................................... 19 1.3.3. 1.4. Earlier literature on the VCSP ............................................................................. 19 1.3.3.1. Partial integration ......................................................................................... 20 1.3.3.2. Complete integration .................................................................................... 21 Combining routing and scheduling: the Vehicle and Crew Routing and Scheduling Problem (VCRSP) ................................................................................................................. 22 2. Present.............................................................................................................................. 23 2.1. Categorization of recent VCSP approaches ............................................................... 23 2.1.1. Categorization according to practical problem ..................................................... 24 2.1.1.1. Type of transportation problem..................................................................... 24 IV 2.1.1.2. Mode of transportation ................................................................................. 25 2.1.1.3. Number of depots......................................................................................... 25 2.1.1.4. Objectives .................................................................................................... 25 2.1.1.5. Size/practicality of the problem..................................................................... 26 2.1.1.6. Degree of urbanization ................................................................................. 27 2.1.1.7. Regularity of the timetable ............................................................................ 27 2.1.1.8. Admission of changeovers ........................................................................... 28 2.1.2. Categorization according to solution method ....................................................... 28 2.1.2.1. Degree of integration .................................................................................... 28 2.1.2.2. Degree of network segmentation.................................................................. 29 2.1.2.3. Model ........................................................................................................... 29 2.1.2.4. Algorithm ...................................................................................................... 30 2.1.2.5. Dynamism of the solution approach ............................................................. 30 2.1.3. Actual categorization ........................................................................................... 31 2.1.4. Conclusions ........................................................................................................ 38 2.2. Evolution of the VCRSP ............................................................................................. 49 2.2.1. Overview of existing approaches ......................................................................... 50 2.2.2. Conclusions ........................................................................................................ 52 3. Future research ................................................................................................................. 57 4. General conclusion............................................................................................................ 65 Bibliography .............................................................................................................................. IX Appendix A .............................................................................................................................. A.1 Appendix B .............................................................................................................................. B.1 Appendix C............................................................................................................................. C.1 Appendix D............................................................................................................................. D.1 Appendix E .............................................................................................................................. E.1 Appendix F .............................................................................................................................. F.1 Nederlandse samenvatting ................................................................................................... SV.1 V Abbreviations We provide an overview of the abbreviations which will been used throughout the thesis. In alphabetical order: CBN Connection-Based Network CG Column Generation CSP Crew Scheduling Problem CVRP Capacitated Vehicle Routing Problem EA Evolutionary Algorithm GDA Great Deluge Algorithm GRASP Greedy Randomized Adaptive Search Procedure ICSP Independent Crew Scheduling Problem LNS Large Neighborhood Search MDVCSP Multiple Depot Vehicle and Crew Scheduling Problem MDVRP Multiple Depot Vehicle Routing Problem MDVSP Multiple Depot Vehicle Scheduling Problem MIP Mixed-Integer Programming PAP Personnel Assignment Problem RCSP Resource Constrained Shortest Path RFS Road Feeder Service RRT Record-to-Record Travel SA Simulated Annealing SDVCSP Single Depot Vehicle and Crew Scheduling Problem SDVRP Single Depot Vehicle Routing Problem SDVSP Single Depot Vehicle Scheduling Problem TS Tabu Search TSN Time-Space Network VCRSP Vehicle and Crew Routing and Scheduling Problem VCSP Vehicle and Crew Scheduling Problem VCSPTW Vehicle and Crew Scheduling Problem with Time Windows VRP Vehicle Routing Problem VRPPD Vehicle Routing Problem with Pickup and Delivery VRPTW Vehicle Routing Problem with Time Windows VSP Vehicle Scheduling Problem VI Definitions We provide an overview of the most important terms, also introduced by (Huisman, 2004) and (Steinzen, 2007), which will been used throughout the thesis. In alphabetical order: changeover change of vehicle of a crew compatible trips trips that can be executed consecutively by the same vehicle deadhead (trip) vehicle movement without carrying passengers or products, such as movements from a depot to the first trip and from the last trip to a depot or idle times outside the depot depot maintenance and storage facility for vehicles that are not in use for some time duty sequence of pieces of work that can be assigned to the same crew piece (of work) sequence of tasks without a break that can be performed without interruption by a single crew staying with the same vehicle relief point location and time where a crew may change vehicles task part of a vehicle block between two consecutive relief points that can be assigned to a crew timetable item that defines the set of trips that are used to carry passengers or deliver products trip vehicle movement with passengers or products specified by start and end locations and times vehicle block sequence of compatible trips and deadheads, starting and ending in the depot, that can be executed by a single vehicle vehicle type a set of vehicles with the same capacity, speed and equipment VII List of figures Figure 1: Planning process in transportation companies p. 12 Figure 2: Planning process in transportation companies (extended) p. 16 List of tables Table 1: Summarized categorization of recent VCSP approaches p. 33-37 Table 2: Two-dimensional classification of relevant VCRSP papers p. 51 VIII 0. Introduction Vehicle routing and crew scheduling are perhaps the most important problems faced in today’s transport companies, since vehicles and especially crews are the main resources for providing services to their customers. This thesis actually covers the whole of the planning process in transportation, comprising both routing and scheduling, more particular the Vehicle Routing Problem (VRP), the Vehicle Scheduling Problem (VSP) and the Crew Scheduling Problem (CSP)1. Slightly bluntly, we could say that the VRP constructs routes so that a number of customers can be serviced with a fleet of vehicles, while the VSP assigns vehicles to cover these routes and the CSP allocates crews to operate the vehicles (and routes). For every one of these planning problems the main objective is to minimize the total costs, taking into account certain constraints. Foregoing definitions of course do not show the total scope and complexity of the subjects, which we will try to clearly and completely describe throughout this thesis. A very important aspect in this regard is the consideration of the possibilities for integration, within the scheduling, so between VSP and CSP leading to an integrated Vehicle and Crew Scheduling Problem (VCSP)2, as well as on a higher level, that is between routing and scheduling and thus defining a Vehicle and Crew Routing and Scheduling Problem (VCRSP). The title of this thesis, Linking the Vehicle Routing Problem and the Crew Scheduling Problem: Past, Present and Future Research, does in fact reflect every one of these aspects of the transportation planning process. The title talks about the linking of the VRP and the CSP, without mentioning the VSP. Yet, this is not an incompleteness. Namely, when using a traditional sequential scheduling approach, the VSP always has to be solved first before being able to start with the CSP, which uses the results of the VSP. In other words, solving the CSP actually implies solving the VSP, so that the title is indeed not incomplete. Because of this incorporation of the VSP in the CSP, it is a logical proceeding to also discuss the VCSP. Moreover, the linking of the VRP and the CSP can be understood as the integration between routing and scheduling and thus announces the discussion of the VCRSP. However, the VCRSP does not emerge in every type of transportation environment, but only in a distribution context and therefore not in public transport. For a first general understanding, public transport companies can be either bus, tram, metro, train or airline operators which are concerned with transporting passengers, whereas distribution companies (e.g. freight transport, 1 These and other abbreviations will be used throughout the entire thesis, mainly with the aim of not overcharging the text. We refer to the introductory section Abbreviations. The reader should also take a good look at the other introductory section Definitions, because some of the terms described there will be used in the initial part of the thesis, before the definition is given in the text. 2 Simultaneous Vehicle and Crew Scheduling Problem or just Vehicle and Crew Scheduling Problem are synonyms for integrated Vehicle and Crew Scheduling Problem. 1 delivery services, mail distribution) deliver products to their customers. The distinction between public transport problems and distribution problems will be an important aspect throughout the entire thesis. To the best of our knowledge, we are the first to make clear use of this natural division and thus also to indicate the impact it has on the general approach of a particular problem. Other papers do mention the existence of both contexts, but do not explicitly point to the implications for a specific problem situated in one or the other context. Of course, the above is not the only contribution of this thesis. In the first place, the thesis comprises a general overview of all planning problems which can arise in a transportation company and discusses them individually and well structured, defining links between the problems afterwards. This is different to other papers that nearly always focus on a single problem like the VRP or the VCSP, which complicates the situating within the bigger picture of the entire planning process. We do provide such a general situating. A well structured overview does indeed appear useful, because it is not unthinkable for a not so experienced reader to no longer see the wood for the trees when reading many publications all discussing specific but somehow connected problems. Own experiences only confirm these findings. We therefore try to provide some sort of guide for reading transportation planning literature. Moreover, through the use of an overall view, interesting and still unidentified subject for future research could be discovered. And even more importantly, it enables us to find a conclusive answer to the central research question of this thesis: ‘Is it (always) advantageous to use an integrated approach for planning problems in transportation?’ There are two aspects to this central research question, being (1) the integration of vehicle and crew scheduling and (2) the integration of routing and scheduling. In other words, we try to determine whether or not the VCSP and the VCRSP, respectively, are beneficial compared to the traditional approaches. To the best of our knowledge, this is one of the first papers – if not the first – to try to answer these questions in a very general and objective way, considering multiple relevant publications all together which is very different from conventional papers in which only the self-proposed method is evaluated. Our goal is to find a more generalized and conclusive answer about the potential of integration as a methodology. Maybe most importantly, this thesis goes further than the mere listing of existing papers. It also does something with it. To the best of our knowledge, it is the first work to introduce an extensive categorization procedure for VCSP approaches. This provides a much needed orderly overview for a quite highly dispersed research area facing a recent booming of publications. The reader (e.g. a planner of a transport company) should thus be able to find the most relevant approach for the particular scheduling problem he is dealing with, among all the available VCSP publications. The proposed categorization is also of use for purely research oriented issues. It can be the basis for identifying gaps within the research area, in that way providing a new 2 perspective on the opportunities and needs for future research. Although not yet extensively elaborated, an impetus for categorizing VCRSP approaches is also given. It is beyond the scope of this thesis to explicitly present mathematical formulas and thus fully describing the proposed models and algorithms. For this we always refer to the concerning papers. A clear but brief description of the most important models and algorithms will be provided though. However, this will never be detailed enough to be able to make propositions for the adaptation of specific models or algorithms. It is also not our objective to identify the ‘best’ method to tackle a certain problem. We confine ourselves to mentioning which methods are ‘good’. Another limitation concerns the maximum allowed volume of this thesis. Since we attempt to give a broad overview of the planning process in transportation, not all aspects can be extensively treated. We especially limit the literature review for non-integrated problems and less recent integrated approaches, for obvious reasons. Moreover, some less essential parts of methods and definitions will not be presented in the main body. Those are added as appendices, mostly literally quoting from existing papers. The imposed time limitation (together with the volume restriction) made it impossible to deeply delve into all possible aspects of the planning problem. Nevertheless, the presented quotes are selected to perfectly complement3 the discussion held. The organization of the thesis can be inferred from the title, for we introduce the chapters Past, Present and Future research. For several reasons, the year 2004 was imposed as the boundary between Past and Present. Every chapter is further subdivided into sections. At the end, we summarize the most important elements and conclusions of this thesis in the closing chapter General conclusion. We now briefly describe the content of the chapters composing the main body. In the chapter Past we first discuss the individual routing and scheduling problems, so the VRP, the VSP and the CSP. We then situate these individual problems within the larger whole of the planning process in transportation companies, also elaborating on the distinction between public transport and distribution context. Afterwards, the scheduling problem is described, first the traditional sequential approach (VSP first - CSP second) and then the integrated VCSP approach, defined in both public transport and distribution context. Subsequently, the most 3 The term complement indicates that the thesis can be understood without the reading of the appendices, which are thus not an inherent part of the thesis, however providing useful extra information and insights. 3 important earlier (pre-2004) literature on the VCSP is briefly reviewed4. To close the chapter, the concept of combining routing and scheduling, thus the VCRSP, is introduced. The chapter Present starts with the most important part, being the categorization of recent VCSP approaches. Of course, we first define the categorization criteria, which either relate to the practical problem considered or the solution method used for it. Once this is done, we can proceed to the actual categorization of the VCSP approaches. Next, conclusions are derived from the presented categorization, including an answer to the first aspect of the central research question. The second part of the chapter is dedicated to the VCRSP, beginning with an overview of the existing approaches and providing an impetus for categorization (after the image of that for the VCSP). Finally, conclusions are stated for the VCRSP, now comprising an answer to the second aspect of the central research question. In the chapter Future research, we orderly list and discuss the research topics deduced from the conclusions for the VCSP and the VCRSP in the previous chapter. The relevance of the topics is substantiated by literal recommendations for future research presented by other authors. Following the logical time scale, we start our exposition with the chapter Past. 4 When literature on a certain problem is presented in this thesis, this will always be done in a chronological manner, so the evolution over time of the can be tracked. If for some reason we should deviate from the chronological ordering, this will always be declared. 4 1. Past In order to place the following discussion into the right perspective, let us start with a definition of what is meant by the term past. Simply stated, the past comprises all the papers that were published before 2004. These papers and the methods they discuss are thus regarded as somewhat older or less recent. Now why did we choose the year 2004 (more precisely January 1st of this year) as the transition point between past and present? There are several reasons for this. As will become clear in section 1.3.3, only a few publications in the pre-2004 operations research literature address a simultaneous approach to vehicle and crew scheduling, which is a very important topic in this thesis and therefore makes a logical criterion for dividing our time line. Moreover, none of those publications makes a comparison between simultaneous and sequential scheduling. Hence, they do not provide any indication of the benefit of a simultaneous approach. This means that the central research question of this thesis (‘Is it (always) advantageous to use an integrated approach for planning problems in transportation?’) was not considered until 2004 ((Huisman, 2004) was the first). And most importantly, the quantity (and quality) of publications on the VCSP increased rapidly since 2004. Of course, we do not only discuss the integration of vehicle and crew scheduling in this thesis, the integration of the routing aspect and the scheduling aspect is also considered. Papers on the latter topic are however even more scarce and the relevant ones date from 2006 (Hollis et al., 2006) at the earliest. Therefore, the choice of 2004 as transition point was deemed logical. This chapter is organized as follows. In section 1.1 we first discuss the individual routing and scheduling problems, so the VRP, the VSP and the CSP. Then, in section 1.2, we situate these individual problems within the larger whole of the planning process in transportation companies, also elaborating on the distinction between public transport and distribution context. In section 1.3, the scheduling problem is described, first the traditional sequential approach (VSP first CSP second) and then the integrated VCSP approach, defined in both public transport and distribution context. We also briefly review the most important earlier (pre-2004) literature on the VCSP. Finally, the concept of combining routing and scheduling, thus the VCRSP, is introduced in section 1.4. 1.1. Individual routing and scheduling problems In this section, we will give a short description of the Vehicle Routing Problem (VRP), the Vehicle Scheduling Problem (VSP) and the Crew Scheduling Problem (CSP) as separate 5 problems. Slightly bluntly, we could say that the VRP constructs routes so that a number of customers can be serviced with a fleet of vehicles, while the VSP assigns vehicles to cover these routes and the CSP allocates crews to operate the vehicles (and routes). For every one of these planning problems the main objective is to minimize the total costs, considering certain constraints. A more accurate description follows in the next subsections, for which the papers of (Toth & Vigo, 2002), (Huisman, 2004) and (Steinzen, 2007) were a great inspiration. In the following section, we describe the three individual problems, also citing some milestone papers. Partially against the structure of this thesis (we are now in the chapter Past), we also cite a few publications from 2004 and later. We do this for the benefit of the completeness of the discussion and because in the chapter Present, the individual problems (VRP, VSP and CSP) will not be re-examined. 1.1.1. The Vehicle Routing Problem (VRP) The goal of the Vehicle Routing Problem (VRP) is to determine the optimal set of routes to be performed by a fleet of vehicles (operated by drivers5) in order to service a given set of customers. (Toth & Vigo, 2002) Because of the diversity that exists within the subject of VRP, this section mainly consists of an overview of the most important types of VRP's. We perform this discussion in the distribution or product delivery context, because this is the context in which the VRP will nearly always occur, unlike in public transport. This will be explained later on in section 1.2. The standard version of the VRP deals with the distribution of a single product from a central depot to a collection of geographically dispersed customers with a deterministic demand, within a single period6 planning horizon. We also call this the standard Single Depot Vehicle Routing Problem (SDVRP). A set of possible vehicle routes for the supply of the customers is designed, taking into account certain constraints. These constraints usually comprise the limited capacity of a vehicle and the customer demand. A VRP that considers the first constraint, is sometimes specified as a Capacitated Vehicle Routing Problem (CVRP). A minimization of the total transport costs is always envisaged. It is implicitly assumed that the supply capacity of the central depot is 5 In this thesis, driver is a general term that not only represent the obvious bus, train and truck drivers, but also pilots, etc. Even stronger, the term driver may also comprise other possible crew members that do not actually drive the vehicle, like conductors (for trains) and flight attendants (for airlines). In that case, driver can in fact be seen as a synonym for crew. Nevertheless, the term driver is often intuitively used (just like here) because most of the time concepts are initially presented for the most widely discussed mode of transport, namely road traffic (where generally there is only one crew member – the driver – per vehicle). 6 This means that a particular time frame, usually a period of one day, is respected as a planning horizon. 6 sufficient to meet the needs of all customers and that each customer can be reached from the central depot by a roundtrip on the same day. In this way, the standard VRP will eventually strive for the determination of routes which can be ridden by a single driver on a particular day. In the next three paragraphs, we discuss the main subdivisions and variations of the VRP. When the capacity of the supplier is large enough, it may happen that several depots are responsible for the supply of many customers. The problem, which is called Multiple Depot Vehicle Routing Problem (MDVRP), then becomes more extensive as there is a choice from which depot products have to be supplied. In practice, a commonly used technique consists of assigning a certain number of customers to a depot, producing a cluster of customers. It thus becomes sufficient to solve the VRP corresponding to the single depot case for each cluster separately. The distinction between single depot and multiple depot will also prove to be very important for other planning problems discussed later on, because it is a distinction that really determines the nature of the problem and therefore the solution methods used for it. An extension of the VRP which has become more and more popular in recent years because of its applicability to a wide area of real-world problems, is the VRPTW or Vehicle Routing Problem with Time Windows. This problem can be described as a CVRP where each customer defines a certain time period – the time window – in which the pickup or delivery of a product should take place. Remark that the standard VRP does not consider these time windows. The ultimate goal of the VRPTW is to determine routes that begin and end at a depot, serve a certain number of customers en route within the specified time window and ensure that the capacity limitation of the vehicles is not exceeded, all while keeping the routes as short as possible. Not surprisingly, the VRPTW proves to be NP-hard. Another variation of the standard problem is the Vehicle Routing Problem with Pickup and Delivery (VRPPD) in which an amount of goods needs to be transported from certain pickup locations to other delivery locations. The goal is again to find optimal routes for a fleet of vehicles to visit the pickup and drop-off locations. A very important aspect of the VRP is that almost all existing approaches do not distinguish between a vehicle and its driver. A driver is thus identified with the vehicle he drives. This means that the constraints imposed on drivers (work regulations, etc.) are imbedded in those associated with the corresponding vehicles. (Toth & Vigo, 2002) 7 For more information about the VRP and its variants, we refer to (Toth & Vigo, 2002). More recent evolutions in this area of research are discussed in (Golden et al., 2008). 1.1.2. The Vehicle Scheduling Problem (VSP) Vehicle scheduling is the process of assigning trips to vehicles (or vehicles to trips if you will, in any case there has to be a mutual assignment) so that the total vehicle costs are minimal. Generally, it is assumed that start and end locations are fixed for all trips. When in a public transport context, the same can be said for the start and end times of the trips. In a distribution context however, we usually implement time windows so trip times would indeed not be fixed. More information on this subject will follow in section 1.2. As stated in previous section, the distinction between single depot and multiple depot determines the nature of a problem. Therefore, we will approach the discussion of the Vehicle Scheduling Problem (VSP) accordingly. Elements of (Huisman, 2004) and (Steinzen, 2007) were used as a basis and complemented by additional considerations. 1.1.2.1. Single depot case In the case that there is only one depot, a homogeneous fleet and no time constraints, we are dealing with the standard Single Depot Vehicle Scheduling Problem (SDVSP). The standard problem can be defined as finding an assignment of trips to vehicles so that each trip is assigned exactly once, each vehicle performs a feasible sequence of trips (two consecutively assigned trips should be compatible7), each sequence starts and ends at the same depot, and vehicle costs are minimized. The vehicle costs consist of a fixed component for every vehicle (investment and maintenance costs) and variable operational costs for idle and travel time. In most practical situations, companies try to minimize their fixed costs first and leave operational cost minimization as a secondary objective. Usually, it is allowed that a vehicle returns to its own depot between two trips if there is enough time to do this. (Huisman, 2004), (Steinzen, 2007) The SDVSP is not only an interesting problem in itself, it also appears as a subproblem in much more complicated problems like the Multiple Depot Vehicle Scheduling Problem (MDVSP) 7 “If two trips i and j are consecutively assigned to the same vehicle, the start time of trip j should be larger than or equal to the end time of trip i plus the travel time from the end location of trip i to the start location of trip j.” (Huisman, 2004, p. 24) 8 discussed in the next subsection and the integrated Vehicle and Crew Scheduling Problem (VCSP) discussed in section 1.3.2. That is why there has been paid a lot of attention to the subject. For a fairly recent survey on the SDVSP and its practical extensions we refer to (Bunte & Kliewer, 2006). 1.1.2.2. Multiple depot case When there are multiple depots, we are of course dealing with a Multiple Depot Vehicle Scheduling Problem (MDVSP). The transport company now operates its (homogeneous) fleet out of several depots, where each vehicle is associated with a single depot. Again, every trip has to be assigned to exactly one vehicle. The MDVSP can be extended by introducing multiple vehicle types and by the constraint that some trips have to be serviced by vehicles from a certain subset of depots. In some cases there are also other constraints concerning depot capacity or route time, which can be imposed to make instances more realistic and thus more challenging. The main objective remains of course the minimization of the vehicle costs, which can be defined in the same way as in the single depot case. (Huisman, 2004), (Steinzen, 2007) The number of depots proves to be an important aspect defining the complexity of a specific problem instance. Whereas the SDVSP is solvable in polynomial time, (Bertossi et al., 1987) showed that the MDVSP is NP-hard from the moment there are at least two depots. Because the problem is always NP-hard, early works – dating back about 30 years – mainly focused on heuristic algorithms. A fairly recent comparison of different heuristic approaches to the MDVSP can be found in (Pepin et al., 2006). It took until 1989 with (Carpaneto et al., 1989) before exact methods where used to solve the MDVSP. (Fischetti et al., 2001) categorize the exact solution approaches into three basic types, based on the mathematical formulation used: singlecommodity flow, multicommodity flow or set partitioning. An explanation of these formulations will be provided in section 2.1.2.3. (Huisman, 2004) and especially (Steinzen, 2007) provide an elaborate presentation of the three types of exact methods. For a more extensive survey on the specific models for the MDVSP the reader is again referred to (Bunte & Kliewer, 2006). 1.1.3. The Crew Scheduling Problem (CSP) Again, elements of (Huisman, 2004) and (Steinzen, 2007) were used as a basis and complemented by additional considerations. Firstly, we are inclined to state that the Crew Scheduling Problem (CSP) is perhaps the most important planning problem in a transport company since crew costs generally dominate vehicle costs (Bodin et al., 1983). It therefore has 9 received considerable attention in the operations research literature since the late eighties. The problem deals with assigning tasks to crew duties so that each task is performed, each duty is feasible with respect to a set of working rules and the total labor costs are minimized. The working rules may comprise federal laws, safety regulations and (collective) in-house agreements. Examples of working rules are minimum/maximum driving time, maximum spread (length) of the duty, minimum break length and allowed start and end time of a duty. Of course, in practical applications there can be many more of such rules. It is mainly because of this wide variety of duty feasibility constraints that the CSP is more complex than the VSP, whereas in essence both problems are quite congruous. In the basic version of the problem, there are only working rules that are defined on an individual duty, so no rules which require that a minimum/maximum number or percentage of the duties has certain properties are taken into account8. Another basic assumption is that all crews are equal since individual crew members are not considered. This assumption can be extended by introducing multiple crew types. As already mentioned, the main objective is to minimize the labor costs. In practice, this objective is often simplified to minimizing the total number of duties so that only fixed crew costs are necessary to take into account. An extension can consist of also considering minimization of the total working time, which implies the introduction of an hourly rate for working time. On top of that, crew scheduling problems are often subject to non-linear costs like overtime bonuses. (Fischetti et al., 1989) show that in principle every CSP is NP-hard, even when only very basic constraints are considered. (Huisman, 2004), (Steinzen, 2007) The CSP is usually formulated as a set partitioning or set covering problem9. These are usually solved with a column generation approach. Such an approach generally consists of the following (consecutive) components: the mathematical formulation, the master problem, the pricing problem (or column generation subproblem) and the construction of feasible (integer) solutions. In such a model, the working rules in a duty have to be taken into account only in the pricing problem. The master problem can be solved by LP relaxation10 or by Lagrangian relaxation. The first approach generally proves to be the most popular. To get integer solutions, different algorithms, exact or heuristical, can be used. Alternatively to column generation, the 8 “For instance, the percentage of split duties that have two pieces of work – one in the early morning and another in the late afternoon with a long break in the middle – is often restricted.” (Steinzen, 2007, p. 7) 9 The formulation of the CSP as a set covering problem allows tasks to be over-covered, as opposed to the set partitioning formulation. “In practice, this over-covering is often not acceptable, but solutions of this model often contain little or no over-covers (since it is cheaper to assign only one driver to a task). The main advantage of a set covering over a set partitioning formulation is that continuous and integer solutions can be easier computed.” (Steinzen, 2007, p. 36) 10 LP relaxation is a synonym for linear programming relaxation or the shorter linear relaxation. All three terms are used interchangeably throughout the thesis. 10 set of columns can be enumerated or generated heuristically. Furthermore, several metaheuristics have also been proposed for solving the crew scheduling problem. For an overview of the concrete CSP solution methods belonging to one of the above classes, we again refer to (Huisman, 2004) and (Steinzen, 2007). A presentation of even earlier works on crew scheduling can be found in (Carraresi & Gallo, 1984). 1.2. The planning process in transportation companies In this section, we will elaborate on the different phases of the planning process for transportation problems. Now first, why is this planning process so important in the transport sector, especially in public transport? The last few decades, public transport experiences strong competition from private traffic. Especially for environmental and health reasons, it is important to reverse this balance back towards the side of public transport. However, given the difficulty of imposing restrictions or penalties on private drivers, or discounts on the side of public transport, the designated tool for obtaining this reversal is the quality of public transport. It soon becomes clear that a well thought out planning is crucial in this view, as two opposing goals must be satisfied: delivering an attractive service to clients and doing so with as little money as possible. This planning is even more challenging when one realizes that the conflict between quality and cost gives rise to a high degree of e.g. technical requirements (engineering). (Schellinck, 2009) Therefore, the planning process is of the utmost importance. We already described the VRP, VSP and CSP in previous sections, but how do they fit within the greater whole of the planning process of a transport company, in public transport as well as in the distribution context? From the last part of previous sentence, an important subdivision of practical planning problems, which will be explained later in this section, already becomes apparent. For the present, the intuitively obvious distinction between a public transport or mass transit company (which can either be a bus, tram, metro, train or airline operator that transports passengers) and a distribution or product delivery company (e.g. freight transport, delivery services and mail distribution, which deliver goods to particular customers) is sufficient to be able to comprehend the following paragraphs. Note that we include airlines in public transport, while in some papers it is seen as a separate category (Hollis, 2011). Before we answer the opening question of the preceding paragraph, let us clearly distinguish between routing problems (VRP) and scheduling problems (VSP and CSP). If the customers being serviced have no time restrictions and no precedence relations exist, then the problem is 11 a pure routing problem. If there is a specified time for the service to take place (so there is a specified timetable), then a scheduling problem exists. This is the case for public transport problems. Otherwise, we are dealing with a combined routing and scheduling problem. Think for example of a distribution problem where the customers have to receive their order within a given time window. This is not a pure routing problem, because the time windows do impose certain time restrictions. Now we discuss the situating of the VRP, VSP and CSP within the planning process of a public transport company11. Later on, we point out the similarities and differences of this process in a distribution company. The entire planning process in a public transport company is very complex and for this reason not (yet) computationally tractable as a whole. Traditionally, it is therefore divided into three consecutive phases: a strategic, tactical, and operational one. Some phases are further split into several subproblems that have to be sequentially solved. Figure 1 represents the complete planning process. We now concisely describe each phase and the subproblems of which they consist, loosely following (Steinzen, 2007). Figure 1: Planning process in transportation companies 11 The exposition given here is based on the planning for a bus company, so for a road traffic mode of transportation. The planning process for other sorts of public transport companies (train, airline, etc.) is mostly similar, though there are some differences. In section 2.1.1.2 we elaborate on this issue. 12 The strategic phase comprises network design and route planning12. In a normal public transport company, we typically deal with a planning horizon of several years. To start the planning procedure, a so called origin-destination matrix is needed. This matrix actually represents the demand data, that is the number of passengers that want to travel between any two locations in the network on a particular time of the day. During network design we then determine links of sufficient capacity within the network in such a way that the construction costs are kept as low as possible. The subsequent route planning problem then aims at satisfying the passenger demand by selecting a set of line routes and their frequencies, given the transportation network designed in the previous subproblem. Two conflicting goals can be identified, namely the maximization of passenger comfort (measured e.g. by the total transit time or the number of direct connections) and the minimization of line operating costs. From this discussion we can conclude that the strategic phase, especially the route planning phase, largely corresponds to the VRP as described in section 1.1.1. Also the customer-relatedness is an important binding aspect. In the tactical planning phase a timetable is constructed. This timetabling problem is solved on a seasonal basis (e.g. once in winter, once in summer). We presume that line routes, the traveling times along them, their frequencies and any possible layover times at depots or other relief points are known. The actual objective in this phase is to transpose the desired line frequencies into a detailed timetable. Such a timetable defines the start and end locations and times of the set of trips to be performed by the company. Notice that we incorporate the timetabling problem in the tactical phase (following the idea of (Desaulniers & Hickman, 2006)) instead of seeing it as a part of the operational phase as was traditionally done, like e.g. in (Huisman, 2004). Since timetabling has a direct impact on both passengers (or more general customers) and staff, it can be seen as an interface between the strategic and operational phase. As timetabling is an interface, we do not consider it to be an inherent part of either the routing or the scheduling problem. The operational planning is concerned with the construction of vehicle and crew roundtrips with the aim of minimizing total costs while considering a variety of operational constraints and work regulations. One can notice that the operational phase is indeed more focused on the staff (or crew) of the company than on the customers, as was more the case at the strategic level. Now for the different subproblems. First, the Vehicle Scheduling Problem (VSP) consists of assigning vehicles to trips, resulting in a schedule for each vehicle. Such a schedule is then split into 12 The terms line routes or lines are synonyms for routes, but where routes is a more general term, line routes or lines are more appropriate in a public transport context (think of e.g. bus lines). Here we used the more general term of route planning because we will also discuss a distribution context. Although, for a discussion of public transport, this could more fittingly be called line planning. 13 several vehicle blocks, where a new vehicle block starts at each departure from a depot. Every vehicle block is in itself a sequence of tasks. Each of these tasks then needs to be covered by a crew duty (or just duty), which is the mission of the Crew Scheduling Problem (CSP). A duty represents the workload on a daily basis of a not yet specified driver. Of course, a number of work regulations (e.g. sufficient rest time) must be satisfied. From the short-term (daily) duties, the crew rostering problem – which we do not discuss further in this thesis – constructs longterm (monthly) work schedules, logically called crew rosters, again considering multiple work regulations. Unlike duties, crew rosters are assigned to specific drivers. In a distribution company, the general structure of the planning process is more or less similar. The natural sequence of the emerging problems remains the same: VRP, VSP and then CSP. Only the definition of the problems is somewhat different. To indicate this, we will reconsider the three phases in the planning process and point out the similarities and differences with respect to the explanation mentioned for a public transport company. In the distribution context the VRP does of course not consist of designing line routes (e.g. bus lines) that remain the same for a long time and have the aim of picking up passengers at fixed places. In fact, there is no such thing as ‘passengers’ here, it are the products ordered by the customers that have to be transported, not the customers themselves. Those customers are not always known well in advance, nor are their location or their demand (i.e. the demand data constantly changes). Therefore, the routes will not be developed e.g. every few years, but much more frequently, for example with a planning horizon of a few weeks or even less (think of a delivery service that delivers products to customer’s home, within a week from ordering). For the other aspects of strategic planning, we can state the following. Obviously, the goal of the network design problem is still to determine the links of the network in such a way that construction costs are minimized. Meanwhile, the concept of frequency of a chosen route is no longer relevant, because the routes themselves may constantly change. The route planning problem has similar objectives as in public transport: maximizing customer satisfaction and minimizing operating costs of the routes. The objective of customer satisfaction (related to service level) originates from the fact that customers like to get their ordered product on time, which is actually more an operational concern than a strategic one. In the timetabling phase, it is presumed that the routes and their frequencies are known. In the product delivery context however, routes are not fixed and the concept of frequencies is not even relevant. In this respect, we could state that the timetabling problem as such does not arise in a distribution context. At some point though, mostly not long before the actual 14 execution, the routes will be laid out and an exact starting time for the tasks situated within their time windows will be chosen, so actually a ‘timetable’ is indeed created. However, this process is run through so many times – perhaps even every day – that the timetable will always be different. Therefore, we say that timetables are not given in a distribution context. From above discussion, we may certainly conclude that the tactical phase has a whole different nature for distribution companies. Then again, the operational phase in distribution companies is very similar to that in public transport. Still, a difference worth mentioning is that in a distribution context, we also have to consider the imposed time windows during the scheduling phases. The division into public transport and distribution (or product delivery) problems can now be made and is an important classification criterion for the practical planning problems – and for the papers in which they are discussed – considered in this thesis. This classification criterion is quite high-level, because it really suggests a different nature of the planning problem that has to be solved. As already mentioned, public transport problems and distribution problems have quite a few similarities, but then again there are some differences making the problems intrinsically distinct. We will clarify the most important difference through following reasoning. In public transport, the strategic planning horizon is typically several years. So, once designed, the line routes remain unaltered for a long period (e.g. at a particular bus stop, busses are generally expected to pick up and drop off passengers at the exact same time every day). Moreover, the corresponding timetable is adapted only to take seasonal effects into account (e.g. the alternation between school days and vacations influence the frequencies of bus lines). This means that timetables are mostly given for public transport problems. In other words, decisions about which lines to operate and how frequently (i.e. the timetable), are input for the operational planning process in public transport. They can be either determined by the marketing department of the company or imposed by local, regional or national authorities. (Huisman, 2004) This last sentence denotes that public transport companies are not always free to develop their own (optimal) routes, which is generally not the case for distribution companies. Therefore, it is not surprising that the vast majority of papers about planning in public transport do not consider the strategic and tactical phase, but only the operational phase. This means that in most practical public transport problems, only the scheduling aspect and not the routing aspect is considered. Of course, for distribution problems this does not apply. There we have to design new routes much more frequently, so the routing problem is a substantial part of the planning problem that cannot be neglected. Obviously, the scheduling problem must also be 15 tackled in the distribution context. From this perception, we might be tempted to conclude that distribution problems are more difficult than public transport problems because both the routing and scheduling aspect have to be considered. This is however somewhat short-sighted, because – as already mentioned in the beginning of this section – the delicate trade-off between customer and cost is of great importance in public transport so more circumspection and therefore more effort is required. In the preceding description of the general planning process in public transport, we included the strategic and tactical phases. While in practice, these phases are usually not considered. Consequently, we may generally indicate public transport problems and distribution problems as follows in Figure 2. Figure 2: Planning process in transportation companies (extended) 1.3. The scheduling problem Although the routing problem precedes the scheduling problem in the planning process (see Figure 2), we first discuss the latter. This because scheduling has to be performed for every practical transportation problem, in public transport as well as in a distribution context. The routing on the other hand is generally only considered in papers about distribution problems, so 16 this item is somewhat more specific. Moreover, the research and associated literature in the field of scheduling is more extensive and will therefore cover the major part of this thesis. In this section, we actually present the evolution of integration within the scheduling: from VSP first - CSP second to VCSP. The organization is as follows. We first describe the traditional sequential approach (VSP first CSP second) in section 1.3.1 and proceed to the integrated VCSP approach (section 1.3.2), defining it in both a public transport and a distribution context. In the last section 1.3.3 we briefly review the most important earlier (pre-2004) literature on the VCSP, introducing the division into partial and complete integration approaches. 1.3.1. Traditional approach The traditional sequential approach for vehicle and crew scheduling concerns the solving of first the vehicle scheduling problem, and then the crew scheduling problem (in short: VSP first - CSP second). Vehicle schedules are thus determined before crew schedules. In other words, we first assign trips to vehicles and then schedule crews based on the resulting vehicle blocks. In fact, the sequential method for solving a vehicle and crew scheduling problem basically comes down to solving the CSP. Because the solution method for a standard CSP is based on the output of the VSP, the VSP always has to be solved first, which corresponds to the definition of the sequential approach. In Appendix A we present a detailed sequential method following these principles and based on (Huisman, 2004). Of course, other sequential approaches do exist (e.g. with incorporation of time windows when in a distribution context), but the presented one is quite common and provides a basis for several other algorithms, so it will at least give a general understanding of how a sequential method for vehicle and crew scheduling is performed concretely. Looking into the method described in Appendix A may also be useful – though not necessary – as an introduction to the integrated problem introduced in the next section since integrated problems sometimes include traditional sequential vehicle and crew scheduling problems as subproblems. Furthermore, the traditional approach is used as a point of comparison to evaluate the (added) efficiency of integrated methods. 1.3.2. Defining the Vehicle and Crew Scheduling Problem (VCSP) Although the traditional independent scheduling of vehicles and crews as described in previous section was already seriously criticized in the early eighties by (Bodin et al., 1983), long after, most of the algorithms published in literature still followed the sequential approach. Fortunately, 17 it became more and more clear that this was not the right track to follow and thus an evolution towards an integrated approach was needed (an overview of the potential benefits of integration is provided in (Freling et al., 1999)). We repeat that it is in fact so that personnel costs are usually higher than vehicle costs and it would therefore certainly be useful to consider crew scheduling simultaneously with vehicle scheduling. Another argument for integration follows from the straightforward dependence between the two problems. Indeed, it should be understood that the characteristics of the vehicle-related aspects will influence the resulting personnel tasks, while an optimal personnel allocation will surely result in a different vehicle use. It was therefore gradually accepted that an integrated approach of the VSP and the CSP might well lead to significant cost reductions. Thus arose the integrated Vehicle and Crew Scheduling Problem (VCSP). The VCSP arises in both the public transport and the distribution context. In order to incorporate domain specific concepts, we need different descriptions for both contexts. Therefore, we successively and independently of one another discuss the definition of the VCSP, first for the public transport context, and then for the distribution context. There will be overlap to some extent, but we believe that the separate treatment of both problems will benefit the clarity. The discussions are based on (Huisman, 2004) and (Steinzen, 2007), supplemented with additional considerations. 1.3.2.1. Public transport context The VCSP in public transport is the following: “given a set of service requirements or trips within a fixed planning horizon, it minimizes the total sum of vehicle and crew costs such that both the vehicle and the crew schedule are feasible and mutually compatible. Each trip has fixed starting and ending times (so there is a given set of timetabled trips) and can be assigned to a vehicle and a crew member.” (Huisman, 2004, p. 78) In the multiple depot case, we must add ‘from a certain set of depots’ to this last sentence. Of course, if every trip can only be assigned to a vehicle and a crew from one specific depot, the multiple depot problem decomposes to a number of single depot VCSP’s. Furthermore, the travelling times between all pairs of locations are assumed to be known. The VCSP is NP-hard since (at least) the crew scheduling part is NP-hard. When considering multiple depots, we know from section 1.1.2.2 that the vehicle scheduling subproblem is also NP-hard, unlike in the single depot case. The remainder of the definition is included in Appendix B, because it is not per se needed for the further reading of this thesis (in fact, the presented terms in the introductory section Definitions are sufficient and were actually already explained in the discussion of the operational phase of the planning 18 process in section 1.2), although of course very interesting and possibly providing some extra insight into the subject. For example, it should become clear from the extended discussion that a wide variety of practical scheduling problems exist which can be modeled in just as many different ways. In this respect, a categorization of these practical problems and their solution methods would prove very useful to get a well structured overall view on the topic of scheduling in a transportation context. Therefore, this is one of the main goals of this thesis and will be performed in the chapter Present. The exposition is strongly based on (Huisman, 2004), but was slightly reorganized and supplemented in order to make it somewhat completer, mainly to clarify the relation and transition between the single depot and multiple depot case. 1.3.2.2. Distribution context In public transport, the trips (and thus tasks) to be performed by a company are usually defined by a given timetable and have fixed starting times. Therefore, the concept of time windows is not relevant in that context. In the distribution context on the other hand, each task can be characterized by a starting time window, within which the task can be commenced. These time windows originate from the fact that usually there are time windows for deliveries, defined by an earliest pickup time of the product at the depot and a latest drop-off time at the customer. This implements that the routes themselves will have a time window within which they can be operated, seeking to deliver all products on time. This practical application leads to the distribution-specific Vehicle and Crew Scheduling Problem with Time Windows (VCSPTW). The problem is that of creating mutually compatible vehicle and crew schedules that lead to a minimum total cost, where vehicles and crews are based at multiple depots13 and have to cover a given set of tasks with associated starting time windows. A detailed definition following (Hollis, 2011) is again added to Appendix C, mainly for the same reasons as mentioned in previous subsection. This description of the VCSPTW considers not only multiple depots, but also multiple vehicle and crew types, which is the most general case. 1.3.3. Earlier literature on the VCSP Now we have defined the VCSP in previous section, we can proceed to an overview of the associated literature from the past, so prior to 2004. We note that all the papers presented in this section deal with the public transport problem. This because, to the best of our knowledge, 13 For the distribution problem we do not make the distinction between single and multiple depot anymore and only discuss the more general latter case. We do this because we do not want overcharge the text, especially since differences concerning the number of depots are very similar to those for public transport. 19 there is no existing literature before 2004 that proposes a general treatment of the VCSP in the distribution context, where timetables are not given in advance and time windows should be implemented (i.e. the VCSPTW). Given the wider practical significance and the greater difficulty of planning in public transport, this may not be surprising. Although simultaneous vehicle and crew scheduling is was demonstrated to be of significant practical interest by (Bodin et al., 1983), not many approaches of this kind have been proposed in the literature before 2004. They mainly dealt with bus and driver scheduling and fall into the category of either partial or complete integration. We use this last distinction to organize the presentation of the relevant papers. Although some authors (e.g. (Hollis, 2011)) do not see methods for partial integration as a VCSP (according to them, only methods for complete integration – as defined in previous section – may carry that title), we also discuss the partial methods because this outlines the evolution of the integration of vehicle and crew scheduling. 1.3.3.1. Partial integration In the eighties and early nineties, a few authors began to become involved in the VCSP. However, it took some while before problems of considerable size and with multiple depots could be handled with a fully integrated method. Therefore, earlier approaches – up to the late nineties – were based on a heuristic or so called partial integration of vehicle and crew scheduling. Similar to (Freling, 1997), we present two types of partial integration and mention the most relevant papers using them: - a crew first - vehicle second approach: performing crew scheduling while including vehicle considerations, construction of a feasible vehicle schedule is done afterwards Most of the approaches of the first category are based on a heuristic procedure proposed by (Ball et al., 1983). (Tosini & Vercellis, 1988), (Falkner & Ryan, 1992), and (Patrikalakis & Xerocostas, 1992) all based their approaches on this paper. - a vehicle first - crew second approach: performing vehicle scheduling while including crew considerations, construction of a feasible crew schedule is done afterwards 20 The two earliest heuristic approaches of the second type were proposed by (Scott, 1985) and (Darby-Dowman et al., 1988). More recently, (Borndörfer et al., 2002) presented another method belonging to this category. (Haase et al., 2001) and (Freling et al., 2003) provide more extensive overviews of methods that partially integrate vehicle and crew scheduling. 1.3.3.2. Complete integration Previous subsection indicates that only very few partially integrated approaches have been suggested for the VCSP. Since the late nineties however, several researchers started to develop different approaches introducing models and solution techniques based on mathematical programming. The first mathematical formulation for the single depot case was proposed by (Patrikalakis & Xerocostas, 1992). However, their model is computationally intractable and therefore the authors had to resort to a partially integrated solution method, which is why this paper was also mentioned in previous subsection. (Freling et al., 1995) slightly altered this formulation and a couple of years later, (Freling, 1997) proposed the first integrated treatment of vehicle and crew scheduling concerning both model and solution method. His solution approach has inspired multiple later publications (e.g. (Huisman, 2004)). (Haase & Friberg, 1999) proposed another exact algorithm for the single depot VCSP. A few years later, (Haase et al., 2001) introduced an interesting crew scheduling formulation for urban mass transit with side constraints for the vehicles, more specifically a bus counter variable. In the airline world, (Cordeau et al., 2001) and (Klabjan et al., 2002) used a similar approach to integrate aircraft routing (which is equivalent to vehicle scheduling) and crew scheduling. (Valouxis & Housos, 2002) described a VCSP that is actually a CSP since drivers are tied to their vehicle. (Freling et al., 2003) considered the SDVCSP in an urban public transport context with a homogeneous vehicle fleet and two crew types. Note that all papers discussed above focus on the single depot problem. To the best of our knowledge, only one pre-2004 paper deals with integration in the multiple depot case, and that is the one written by (Gaffi & Nonato, 1999). For more extensive discussions of the papers presented above, we can refer the reader to (Huisman, 2004), (Steinzen, 2007) and (Hollis, 2011). 21 1.4. Combining routing and scheduling: the Vehicle and Crew Routing and Scheduling Problem (VCRSP) The combination of the routing problem and the scheduling problem (the latter was discussed in section 1.3) only occurs in the distribution context (e.g. freight distribution, postal organizations and delivery services) and not in public transport, where generally line routes and timetables are fixed, dictated for example by a local authority. It is so that, as could already be suspected from the introduction of section 1.3.3, we could not find any relevant pre-2004 paper on this subject. This means that for this section, we are limited to providing a brief description of the problem and postpone further explanation to section 2.2 in the chapter Present. We get straight to the point by providing a simplified description of the traditional planning process in a distribution company. This process typically involves constructing a set of vehicle routes, then producing vehicle schedules by assigning vehicles to the vehicle routes, and building a set of crew schedules so that all vehicle routes are covered. (Hollis, 2011) For the ease of notation and also to show a certain conformity with the VCSP, we will name a distribution problem where routing and scheduling are combined, a (simultaneous or integrated) Vehicle and Crew Routing and Scheduling Problem (VCRSP). We can provide a formal definition of the VCRSP, which should in fact already be clear from the definition of a distribution problem: the VCRSP concerns a planning problem where the required transports have no given timetable (i.e. distribution context), and where a driver-vehicle combination is not considered an inseparable unit. (Drexl et al., 2011) The latter characteristic stresses the difference with a VRP, where a driver is nearly always identified with the vehicle he drives (recall from section 1.1.1). As a consequence, a change in one route may have effects on the feasibility of other routes in a VCRSP (this is called the interdependence problem), which is not the case in a classic VRP where independence of routes is implicitly assumed. Thus we have discovered an additional difficulty of the VCRSP. From this discussion, it should become apparent that a VCRSP consists of some sort of merger between a VRP and a VCSP. Now let us get a little ahead of things and already broach the subject of integration in the distribution context. If in previous section about the scheduling problem, the integration within the scheduling was discussed, then we are now dealing with an integration on a higher level, namely between routing and scheduling (from a traditional routing first - scheduling after approach to simultaneously routing and scheduling). 22 2. Present Since all pre-2004 papers on planning in transportation were said to be part of the past, all publications from 2004 up to now of course belong to the present. The chapter starts with the most important part, being the categorization of recent VCSP approaches. In section 2.1, we first define the categorization criteria, which either relate to the practical problem considered or the solution method used for it. Once this is done, we proceed to the actual categorization of the VCSP approaches. At the end of the section, conclusions are derived from the presented categorization, including an answer to the first aspect of the central research question. The second part of the chapter (section 2.2) is dedicated to the VCRSP, beginning with an overview of the existing approaches and providing an impetus for categorization (after the image of that for the VCSP). Finally, conclusions are stated for the VCRSP, now comprising an answer to the second aspect of the central research question. 2.1. Categorization of recent VCSP approaches Above title already reveals the intent of this section, namely to categorize recent VCSP approaches. Of course, we cannot consider all VCSP approaches introduced as from 2004, but rather select the most important ones. A comprehensive categorization of these approaches (and actually also of the corresponding papers) allows the reader to easily detect e.g. for which practical problem a certain solution method has already been used (and was found to be 'good'), which solution methods are most commonly used for a particular practical problem, which practical problems have been described most frequently, and which solution methods are the most popular. From these observations, we can come to interesting conclusions that may give rise to recommendations for future research (see also chapter Future research). A few possibilities: a solution method that turned out to be very effective for a particular problem may perhaps also be used for other practical problems, some practical problem that did not receive much attention so far – inconsistent with its importance in real-life – can be identified, etc. As could already be suspected from previous listings, we will classify the various VCSP approaches according to two major aspects, namely (1) the practical problem that is considered and (2) the way this problem is dealt with, so the solution method. Within these two overarching categories we define multiple categorization criteria. We identify 8 criteria for the practical problem, and 5 for the solution method. The criteria are presented in descending order of 23 importance (in terms of impact on the nature of the problem14 and its solution method). The proposed criteria are obviously not exhaustive, but we believe to have identified some of the most relevant ones. They are not too detailed, but still comprehensive enough so that all approaches can be categorized sufficiently accurate. This section is organized as follows. We start with the discussion of the categorization criteria within both major categories, again under the pretext of ‘most important first’, so first for the practical problem in section 2.1.1 and then for the solution method in section 2.1.2. Categorization according to practical problem is more important because this actually defines the nature of the problem, something that is in fact fixed and thus an inherent characteristic. Whereas the nature of a given practical problem is unchangeable, the solution method for it can still be chosen to some extent. Of course, the nature of the practical problem will partly determine the solution method (kind of model, etc.) to be used. Therefore, we can state that the categorization according to practical problem is higher level than that according to solution method, which is intuitively logical. In section 2.1.3, we proceed to the actual categorization of the recent VCSP approaches according to the proposed criteria. Finally, we draw conclusions (links and trends) from this categorization, which we can use as a basis for making recommendations for future research. This is included in section 2.1.4. 2.1.1. Categorization according to practical problem The 8 criteria for categorization according to practical problem are (in descending order of importance): type of transportation problem, mode of transportation, number of depots, objectives, size/practicality of the problem, degree of urbanization, regularity of the timetable, and admission of changeovers. For the sake of unity, literal definitions are extracted from one and the same paper of (Steinzen, 2007). 2.1.1.1. Type of transportation problem The first and most high level criterion is one that was already frequently mentioned and discussed thoroughly in section 1.2, namely the subdivision in public transport problems and distribution problems. We will not resume the entire discussion here, but rather refer to the mentioned section. 14 W hen we are simply talking about a problem in this section, we obviously refer to a VCSP. 24 2.1.1.2. Mode of transportation We can identify three major modes of transportation: road traffic, railway and airline. Under the title of railway we include trains, trams and metros. Clearly, this subdivision mainly concerns public transport. Although one can imagine freight trains and cargo planes being used in the distribution context, the vast majority of the distribution problems deal with transport by road. The trucks, lorries and delivery vans used for this are categorized under road traffic, just like the busses used in public transport. The distinction in terms of mode of transportation is an important aspect regarding the nature of – and therefore the solution method for – a scheduling problem, especially its complexity. We therefore include an elaborate discussion of the differences between bus, railway and airline planning (as was promised in section 1.2) in Appendix D. This discussion – inspired by (Huisman, 2004) – is certainly interesting, but not necessary for the further reading of this thesis. 2.1.1.3. Number of depots This criterion does not refer to the exact number of depots considered, but makes the distinction between the single depot case and the multiple depot case. This distinction has been discussed repeatedly in preceding sections and is therefore not repeated here. However, we do recall that the number of depots, also the exact15 number, will have a significant impact on the complexity of the problem. 2.1.1.4. Objectives Needless to say that for any scheduling problem, (1) cost reduction (minimization of costs) is nearly always the main goal. However, there are also other objectives that can be pursued. We will consider two important additional goals. A first one is (2) the optimization of the service level, which concretely means that delays are to be minimized. This objective is more relevant in a public transport context because timetables are very strict here, whereas for distribution problems there is some tolerance due to the 15 By which we mean that, although two problems considering e.g. 2 and 4 depots both concern a multiple depot case (and are thus the same according to the ‘number of depots’-criterion), the exact number of depots does of course have an impact on complexity, that is, the problem considering 4 depots will be more complex and thus harder to solve. 25 existence of time windows (but still, customers of a distribution company want to get their goods on time so service level can also be relevant here). A second additional objective is related to (3) the quality of crew schedules. We consider the regularity of crew schedules as the defining quality aspect. A crew schedule is called regular if it can be repeated many times. “Regularity is an important aspect for crew schedules since regular solutions can improve operational reliability and reduce training costs. Furthermore, regular solutions are less error-prone and crews often prefer to repeat itineraries.” (Steinzen, 2007, p. 151) Generally, regular crew schedules are much more difficult to obtain in a distribution context than in a public transport context because of the ever changing timetables in the former case. We do not consider the regularity of vehicle schedules as a relevant objective. Indeed, vehicles are rather insensitive to the quality of their schedules (as opposed to crew members). Notice that objective (2) is aimed at enhancing the quality for customers, while (3) tries to improve the quality for the crew (and the company). 2.1.1.5. Size/practicality of the problem It is difficult to unambiguously categorize the considered problem in terms of size (i.e. number of trips, depots, vehicle types, etc.) as being small, medium-sized or large. Where do we set the boundaries and how do we deal with the fact that these will shift over time? Because of these limitations, it seems recommended to evaluate the size of a problem in terms of practicality. We therefore classify the problems (and with them the papers in which they are discussed) as either theoretical or application-based. In this thesis, the first class of theoretical problems deals with randomly generated test instances. Mostly, the constraints considered in such a problem, e.g. with respect to the feasibility of a duty, are kept rather simple. In the second class, the proposed algorithms are applied to solve problems which arise from real-life applications and thus contain many complex practical constraints. We could state that theoretical problems generally correspond to rather small problems (i.e. ‘easy’ randomly generated test instances), whereas application-based problems correspond to medium-sized and large problems (i.e. based on data from real-life transport companies). Of course, one can think of a paper in which a proposed method is tested on randomly generated instances as well as on real-life problems. In that case, the application-based classification dominates the theoretical classification. 26 2.1.1.6. Degree of urbanization We immediately start with a definition: “Public transport scenarios can be categorized according to the structure of the underlying transportation network. Urban service provides connections within the city while ex-urban16 (regional) service connects the city with the suburbs and minor towns in the region of the city.” (Steinzen, 2007, p. 152) Although one could notice a difference in gradation, the terms suburban17 and ex-urban are assumed to be equivalent here. Ex-urban services are thus defined as all services that are not strictly urban, which seems to be common practice in transportation literature. Of course, many companies offer a mixture of both urban and ex-urban services. Similar general statements apply in the distribution context. Above explanation is of course suitable for road traffic. For airlines, however, the (ex-)urban categorization is not very relevant, because the distances traveled here far exceed the dimensions of cities. For rail traffic, we could make the analogy by stating that trains operate within an ex-urban scenario while trams and especially metros have a more urban character. 2.1.1.7. Regularity of the timetable A regular timetable is a timetable that remains the same for each day of the week, every week. “In practice however, timetables may consist of many trips serviced every day and some exceptions that do not repeat daily. In particular, service trips to schools, production facilities, or public swimming baths are often subject to change, e.g., trips may be operated on every day except Sunday or on Monday only.” (Steinzen, 2007, p.11) This results in irregular timetables. “Notice that public transport companies face a similar situation whenever they change their timetable, e.g., scheduled timetable changes in summer or winter. Typically, these changes involve only a small portion of the complete timetable” (Steinzen, 2007, p. 12), as is usually the case for irregular timetables. We remark that regularity of the timetable should be seen in the context of presented regularity of the timetable. For example, a certain transportation company can have an irregular timetable in reality (e.g. different scenarios for weekdays and Sundays, i.e. more trips on weekdays), but when this irregularity is not taken into account in the paper describing the real-life instances (e.g. only the weekday scenario is mentioned, so the reader cannot know of the (non-)existence 16 Ex-urban is a synonym for extra-urban, but throughout this thesis we will keep using the term ex-urban. Suburban settings have more relief opportunities and smaller distances between the depots than exurban settings. 17 27 of a Sunday scenario), then the timetable is classified as regular. When, for example, both weekday and Sunday scenarios are presented in the paper, we deal with an irregular timetable. The exposition above primarily relates to the public transport context. In the distribution context, where the timetable is not given, there are many trips – usually even the vast majority – that do not repeat daily. We can therefore state that, for distribution problems, we will nearly always have to deal with (highly) irregular timetables. 2.1.1.8. Admission of changeovers A changeover is the change of vehicle of a driver whenever there is a relief point. This means that, when changeovers are admitted, a crew is allowed to work on more than one vehicle during a duty. We can deal separately with the situation where changeovers are and are not allowed. The first category can be further divided, that is, the admission of changeovers can be restricted or unrestricted. Restricted changeovers may imply assuming that “each crew is assigned to a depot and may only conduct tasks on vehicles from this particular depot (which is of course only a relevant criterion for multiple depot problems) and furthermore, assuming that a driver may only change his vehicle during a break, i.e., between two pieces of work. In other words, a restricted changeover is only allowed between vehicles from the same depot and the driver must take a break after leaving his vehicle. […] However, when changeovers are unrestricted, a driver may change between two vehicles of different depots whenever there is a relief point (no matter if he takes a break or not).” (Steinzen, 2007, p. 110) 2.1.2. Categorization according to solution method The 5 criteria for categorization according to solution method are (in descending order of importance): degree of integration, degree of network segmentation, model, algorithm, and dynamism of the solution approach. 2.1.2.1. Degree of integration For the categorization according to solution method, the most high level criterion is that concerning the degree of integration between vehicle and crew scheduling. As already mentioned in section 1.3.3, we can distinguish between methods for partial integration and methods for complete integration. The former can be further subdivided into crew first 28 vehicle second and vehicle first - crew second approaches. For more information about this subdivision, we refer back to section 1.3.3.1. 2.1.2.2. Degree of network segmentation To solve the scheduling problem, some transportation companies consider each route separately (e.g. each bus line), while others consider a part of the network or even the whole network at once. We remark that this categorization criterion is only relevant when the proposed method is tested on an application-based (real-life) problem. This because theoretical randomly generated instances are almost always tailored in such a way that they can be solved as a whole. 2.1.2.3. Model A model defines the mathematical formulation of a practical problem. Generally, a model can be classified according to the type of formulation it uses: single-commodity flow, multicommodity flow, set partitioning or set covering. It is certainly possible that more than one type of formulation is used in a model. We will give a very limited description of these formulation principles. For a more detailed discussion we refer to e.g. (Toth and Vigo, 2002) and (Huisman, 2004). The single-commodity flow problem (also called minimum cost flow problem) consists of “determining a least cost shipment of a commodity through a network that will satisfy the flow demands at certain nodes from available supplies at other nodes. [...] The multicommodity flow problem is an extension of the single-commodity flow problem, since instead of a single commodity several commodities use the same underlying network. The different commodities now have different origins and destinations.” (Huisman, 2004, pp. 10-11) Given a set of elements (called the universe) and a number of sets whose union comprises the universe, a set covering problem aims to identify the smallest number of sets whose union still contains all elements in the universe. Note that one or more customers may be visited more than once here. In the set partitioning problem this is not the case, because it is imposed that each customer must be covered by exactly one of the selected routes. The four above formulation types are the most common ones, though we mention another one that is also used in rare cases, namely the set packing formulation. In the set packing problem, 29 given a certain universe and a family of subsets of that universe, the task is to find a subfamily that uses the largest number of pairwise disjoint sets. Still other types of formulations are possible, but are only used sporadically. If such a formulation occurs in the actual categorization, we will give a brief description of it there. 2.1.2.4. Algorithm Basically, an algorithm defines the way of solving a practical problem that was converted into mathematical terms forming a model (see previous subsection). A large portion of the VCSP algorithms rely on column generation. Therefore, we present the general structure of such an approach. We can divide a column generation-based algorithm into following consecutive stages: the master problem, the column generation subproblem (or pricing problem) and the construction of feasible solutions. We remark that these components were already mentioned in section 1.1.3 about the CSP. For each of the three stages, multiple solution techniques are possible. A general understanding of the column generation method, together with the presentation of more specific techniques that can be used for the different stages, will follow from the actual categorization in section 2.1.3. When dealing with a VCSP approach based on column generation in the actual categorization, we will always try to describe the specific techniques used for all three components in an orderly fashion. Of course, other algorithms do exist that are not based on column generation (e.g. metaheuristics). These will be described just as well, although the three stages applicable for column generation will obviously not occur here. 2.1.2.5. Dynamism of the solution approach “Traditionally, a VCSP is solved only once, some time (this varies from a few days or weeks in the distribution context to a few months in the public transport context) before the new timetable starts, and it will not be changed for the whole period that the timetable is valid.” (Huisman, 2004, p. 111) Furthermore, travel times are assumed to be fixed. This is called the static approach for solving a scheduling problem. The basic idea of a dynamic approach is that schedules are constructed several times a day. This means that we have to reschedule a few times during the day, reassigning vehicles and crews to trips. Moreover, we can take into account different scenarios for future travel times. 30 (Huisman, 2004) In some papers (see (Barnhart & Laporte, 2007)) the dynamic approach to scheduling is reflected in a real-time control phase (succeeding the three planning phases defined in section 1.2) in which the whole process is evaluated, adjusted and maintained. As one can imagine, a dynamic approach is much more labor intensive than a traditional static one. 2.1.3. Actual categorization Having defined the relevant criteria, performing the actual categorization of the VCSP approaches is now fairly straightforward. Yet, we should make some remarks about the concrete form the categorization will take. First of all, each approach is logically denoted by the paper in which it is presented. When multiple approaches are contained in one paper, a particular approach will be indicated by the extra mentioning of the section in which it is described (e.g. (Huisman, 2004), Ch. 3.3). We also notice that it is not uncommon that quite a few papers actually present virtually the same approach. In that case, we will not categorize all those papers individually, but only perform the categorization for the most relevant one (i.e. the most recent or most comprehensive paper) and mention the other ones as ‘related papers’. In order to provide an orderly overview, the discussion will be performed chronologically. Of course, all criteria for practical problem and solution method will be evaluated. When the categorization according to a certain criterion is not straightforwardly obvious, we will always add a word of explanation. This is certainly useful when describing the algorithm, which is usually not unambiguously definable in a few words, let alone one specific term (as opposed to the model). Therefore, the algorithm will always be described somewhat more extensively. Concerning the criterion size/practicality of a problem, we always name the specific company from which real-life data is used when discussing an application-based approach (paper). To provide some sort of numerical notion about the size of the treated – and thus treatable – problems using a particular method, we also mention the maximum number of trips, depots and vehicle types considered. The discussion of an approach always ends with a conclusion mainly concerning the performance of the proposed method (i.e. a summary of the computational results). The most important remark is still to come. Although all three modes of transport – road traffic, railway and airline – were mentioned in the corresponding criterion, we will mainly focus our categorization on the road traffic case. This has multiple reasons. First of all, it is quite simply so that most relevant publications on planning in transportation deal with road traffic (mainly busses), partially because a larger benefit of integration can be 31 obtained here18. This also explains why we almost always based our discussion of a certain problem or method on the road traffic case. Also the definitions of the categorization criteria are primarily focused on road traffic. For example, the classification urban/ex-urban is very relevant for road traffic, but not at all for airlines. However, for the railway case we can use the exact same categorization criteria presented in previous sections 2.1.1 and 2.1.219. We will prove this by categorizing an interesting approach for the VCSP in a railway context, leaving the categorization criteria unaltered. Moreover, recent railway publications are far less numerous, so many more relevant papers were not to be found anyway. On the other hand, terms used in airline planning – and therefore in the categorization criteria for airline problems – do not always entirely correspond to the terms used for road traffic (and railway), although there certainly are some parallels to be discovered. For example, the classification short-haul flights/long-haul flights can be seen as an equivalent for urban/ex-urban and it was already mentioned that the process of aircraft routing is similar to vehicle scheduling (see section 1.3.3.2). As regards solution methods, there are also some similarities to be identified between airline and road traffic, e.g. the model introduced in (Sandhu & Klabjan, 2007) is very similar to the models used in (Hollis, 2011). The presented categorization procedure for VCSP’s can thus fairly easily be extended to cope with airline problems. Nevertheless, we do not explicitly do this ourselves, mainly because of the imposed limitations in size of this thesis. Thus, an issue for future research is discovered. The actual categorization is now performed for 18 of the most important VCSP approaches. For the purpose of not overcharging the main body, the detailed categorization tables are included in Appendix E. We immediately present the summarizing Table 1 here. First, a few comments on this table. Related papers are mentioned (in italics) below the titles of the discussed approaches. For the sake of clarity, the extra explanation of the criteria has been omitted. Sometimes there will be two columns for one approach when listing maximum number of trips, depots and vehicle types. This occurs when the three characteristics are maximum for different instances. Most importantly, the algorithms are described only very briefly, using no more than a few summarizing terms. Nevertheless, these terms should be sufficient to provide a clear view on which type of algorithm was used (a column generation-based method, a metaheuristic, etc.), so comparison on a general level is possible. The meaning of the abbreviations used in the discussion of the algorithms can be found in the introductory section Abbreviations or in the comprehensive categorization of the concerning approach in Appendix E. 18 This can be explained by the fact that the relative difference between crew and vehicle costs is higher for road traffic than for the other modes of transportation. (Huisman, 2004) 19 Although the vehicle scheduling part of the algorithm is much more complicated for railways, but this does not affect the categorization. 32 practical problem transportation problem type mode of transportation number of depots objectives (Huisman, 2004), Ch. 3.3 (Huisman, 2004), Ch. 3.4 (Huisman, 2004), Ch. 4.2-4.3 (Huisman et al., 2005) (Huisman, 2004), Ch. 4.4 (Huisman et al., 2005) public transport public transport public transport public transport road traffic road traffic road traffic road traffic single depot single depot multiple depot multiple depot cost reduction cost reduction cost reduction cost reduction size/practicality of application-based application-based application-based application-based the problem max trips 259 259 653 653 max depots 1 1 4 (avg 1.71) 4 (avg 1.71) max vehicle types 1 1 1 1 degree of urban urban ex-urban ex-urban urbanization regularity of the regular timetable regular timetable regular timetable regular timetable timetable admission of restricted restricted restricted no changeovers changeovers changeovers changeovers changeovers solution method degree of complete complete complete complete integration integration integration integration integration degree of network each route each route part of the part of the segmentation separately separately network network model set partitioning set covering set partitioning set partitioning algorithm dynamism of the solution approach Lagrangian relaxation, column generation, Lagrangian heuristic Lagrangian relaxation, column generation, Lagrangian heuristic Lagrangian relaxation, column generation, Lagrangian heuristic Lagrangian relaxation, column generation, Lagrangian heuristic static approach static approach static approach static approach Table 1 a: Summarized categorization of recent VCSP approaches 33 practical problem transportation problem type mode of transportation number of depots objectives (Huisman, 2004), Ch. 5.3 (Huisman & Wagelmans, 2006) (Borndörfer et al., 2004) (Rodrigues et al., 2006) (Steinzen, 2007), Ch. 2.3-2.4 (Gintner, 2007) public transport public transport public transport public transport road traffic road traffic road traffic road traffic multiple depot multiple depot single depot multiple depot cost reduction and service level cost reduction cost reduction cost reduction size/practicality of application-based application-based application-based application-based the problem max trips 304 1,414 634 395 653 max depots 4 (avg 1.71) 1 3 1 4 max vehicle types 1 3 5 1 1 degree of ex-urban urban/ex-urban urban ex-urban urbanization regularity of the irregular regular timetable regular timetable regular timetable timetable timetable admission of restricted restricted restricted no changeovers changeovers changeovers changeovers changeovers solution method degree of complete complete complete complete integration integration integration integration integration degree of network part of the whole network at each route part of the segmentation network once separately network model multicommodity multicommodity set covering/ set partitioning flow and set flow and set set packing partitioning partitioning algorithm Lagrangian Lagrangian Lagrangian hybrid algorithm, relaxation, relaxation, relaxation, mathematical column column column programming generation, generation, generation, and (greedy) Lagrangian branch-andLagrangian heuristic heuristic bound heuristic dynamism of the dynamic static approach static approach static approach solution approach approach Table 1 b: Summarized categorization of recent VCSP approaches 34 practical problem transportation problem type mode of transportation number of depots (Steinzen, 2007), Ch. 3.1-3.3 (Steinzen et al., 2010) (Steinzen, 2007), Ch. 3.4 (Steinzen, 2007), Ch. 4 (Steinzen, 2007), Ch. 6 (Steinzen et al., 2009) public transport public transport public transport public transport road traffic road traffic road traffic road traffic multiple depot multiple depot multiple depot single depot cost reduction cost reduction and quality of crew schedules theoretical application-based 200 4 1 433 1 1 ex-urban ex-urban objectives cost reduction cost reduction size/practicality of application-based application-based the problem max trips 653 653 max depots 4 4 max vehicle types 1 1 degree of ex-urban ex-urban urbanization regularity of the regular timetable regular timetable timetable admission of restricted unrestricted changeovers changeovers changeovers solution method degree of complete complete integration integration integration degree of network segmentation model algorithm dynamism of the solution approach part of the network multicommodity flow and set partitioning Lagrangian relaxation, column generation, heuristic branchand-price part of the network multicommodity flow and set partitioning Lagrangian relaxation, column generation, heuristic branchand-price static approach static approach regular timetable restricted changeovers irregular timetable restricted changeovers complete integration partial integration (crew - vehicle) n/a n/a multicommodity flow and set partitioning set covering hybrid algorithm, mathematical programming and EA Lagrangian relaxation, column generation, local and followon branching static approach static approach Table 1 c: Summarized categorization of recent VCSP approaches 35 practical problem transportation problem type mode of transportation number of depots objectives size/practicality of the problem max trips max depots max vehicle types degree of urbanization regularity of the timetable admission of changeovers solution method degree of integration degree of network segmentation model (Kéri & Haase, 2007) (Mesquita & Paias, 2008) (Mesquita et al., 2006, 2011) (Laurent & Hao, 2008) (Bartodziej et al., 2009) public transport public transport public transport distribution road traffic road traffic road traffic road traffic single depot multiple depot single depot single depot cost reduction and service level cost reduction cost reduction cost reduction theoretical theoretical 133 1 1 400 4 1 249 1 2 urban ex-urban ex-urban ex-urban irregular timetable regular timetable regular timetable irregular timetable no changeovers unrestricted changeovers no changeovers no changeovers complete integration complete integration complete integration n/a n/a n/a complete integration whole network at once set partitioning multicommodity flow and mixed set partitioning/ covering constraint-based set partitioning application-based application-based algorithm dynamism of the solution approach 1,400 1 n/a 779 1 27 linear relaxation, column generation, round-up method linear relaxation, column generation, branch-andbound metaheuristic, GRASP: constraint programming and local search 1) linear relaxation and column generation 2) local searchbased metaheuristics (SA, GDA, RRT) static approach static approach static approach static approach Table 1 d: Summarized categorization of recent VCSP approaches 36 practical problem transportation problem type mode of transportation number of depots objectives (Sato et al., 2009) (Hollis, 2011), Ch. 4 (Hollis et al., 2006) public transport distribution railway road traffic multiple depot multiple depot cost reduction and service level cost reduction size/practicality of application-based the problem max trips 786 max depots n/a max vehicle types n/a degree of ex-urban urbanization regularity of the regular timetable timetable admission of restricted changeovers changeovers solution method degree of complete integration integration degree of network each route segmentation separately model multicommodity flow algorithm heuristic flow modification, local search dynamism of the solution approach dynamic approach application-based 1,016 23 3 urban/ex-urban irregular timetable restricted changeovers complete integration whole network at once set covering linear relaxation, column generation, branch-andbound static approach Table 1 e: Summarized categorization of recent VCSP approaches 37 2.1.4. Conclusions In this section, we will discuss the trends that can be observed for the VCSP approaches. We will also indicate some links between the presented criteria. Maybe even more important, we try to discover gaps between and within the criteria (which in fact correspond to constraints, e.g. for changeovers) so that we can identify interesting topics for future research that actually consist of filling these gaps. Finally, we are now able to provide an answer to the first aspect of the central research question: ‘Is it (always) advantageous to use an integrated approach for scheduling problems in transportation?’ The order of the proposed criteria for our categorization of VCSP’s (type of transportation problem - mode of transportation - number of depots - objectives - size/practicality of the problem - degree of urbanization - regularity of the timetable - admission of changeovers degree of integration - degree of network segmentation - model - algorithm - dynamism of the solution approach, supplemented by the conclusions for the presented approaches) will be used as a guideline for the division into paragraphs of the upcoming exposition, though it is not strictly followed. It was already mentioned that the great majority of the papers consider a public transport VCSP, as our categorization also confirms (only 2 of the 18 described approaches are situated in a distribution context). However, we cannot draw too many conclusions from this regarding future research (in the sense of ‘the expansion of distribution VCSP literature is needed’) because all papers presented in next section 2.2 do consider the planning problem in the distribution context, just not from a mere scheduling point of view (i.e. the development of a specific VCSP approach) but rather integrating scheduling and routing in one general planning approach. Therefore, suggestions for future research in the distribution context will be provided in next section. It is also remarkable that the distribution papers are two of the most recent ones described, the oldest dating from 2009. So indeed, planning in a distribution context is a very recent research area. Remember that we chose to only consider VCSP approaches for road traffic, supplemented by one paper on railways. The exposition in this section has thus to be understood in the road traffic context. An obvious working point that was already mentioned, concerns the adaptation of the categorization procedure for airline problems and also its expansion in terms of number of considered papers for the other modes of transport, especially railways. 38 Also notice that for railways, it appears from only one considered paper that difficulties like multiple depots, a service level objective and allowed changeovers can be taken into account, just as for road traffic. A small majority of the approaches (11 out of 18) consider the more extensive case of multiple depots. Sometimes authors consider an ‘easier’ single depot case because the addressed practical problem quite simply exists of only one central depot (e.g. the RFS in (Bartodziej et al., 2009) and the RET in (Huisman, 2004), Ch. 3.3 and Ch. 3.4), but mostly they do this to simplify the introduction of a new solution approach (e.g. a hybrid algorithm in (Rodrigues et al., 2006), a method that can deal with irregular timetables in (Steinzen, 2007), Ch. 6, a model that incorporates trip shifting in (Kéri & Haase, 2007) and a novel metaheuristic in (Laurent & Hao, 2008)). An obvious subject for further research is of course to extend these solution approaches to a multiple depot case. Although not generally valid, one could detect a certain relation between number of depots and degree of urbanization. In most cases, an urban context corresponds to a single depot and an ex-urban context to multiple depots. This can be intuitively explained. In an urban context, companies operate within the boundaries of a city, which means trips will never be of great distance so that the vehicles should be able to return to a central depot located somewhere in the city. In an ex-urban context on the other hand, customers are generally much more geographically dispersed so that one single depot might not be able to service all customers (within the imposed regulations), which means that several extra depots would be needed. As already mentioned, this relation is not definite. It is of course possible that an urban transport company chooses to operate from more than one depot, just as it may be that an ex-urban transportation firm is nevertheless able to serve their customers from just one depot. We can state that the service level objective, which aims at reducing delays for customers, actually corresponds to applying a dynamic approach to the problem. Indeed, in both (Huisman, 2004), Ch. 5.3 and (Sato et al., 2009), increasing service level is an objective while a dynamic approach is used to tackle the problem. The conformity of both criteria is in fact quite obvious. When using a traditional static approach, there are no changes in the schedules for the whole period that a timetable is valid. This means that when a delay occurs on a certain day, the next trip to be performed by the concerning vehicle and/or driver may start late, possibly provoking new delays which, in their turn, may cause a similar snowball effect. (Huisman, 2004) Since in a dynamic approach we reschedule a few times per day, we can reassign vehicles and crews to trips and thus may be able to prevent the delays at the start of a trip in many cases, improving the service level for customers. 39 We remark that (Kéri & Haase, 2007) also have service level as an objective although they do not use a dynamic approach. However, we already mentioned in the actual categorization of Appendix E that although not explicitly performed in the paper, it should be logical (and possible – regarding the computation time – for very small instances) to use the proposed approach in a dynamic environment. We attribute the not carrying out of a dynamic approach to the very limited size of the considered paper. Yet, it should certainly be possible to apply the proposed method in a dynamic context, after some minor adaptations. This means the correspondence between the service level objective and the dynamic solution approach is also affirmed by (Kéri & Haase, 2007). The other secondary objective – quality (regularity) of crew schedules – is also related to another criteria, namely that of regularity of the timetable. This should not be surprising, since crew schedule regularity is indeed only a relevant objective when the timetable is irregular (when the timetable is regular, obviously the crew schedules will be too). If a ‘normal’ VCSP approach is used for a problem where the timetable is irregular, regularity of the crew schedules cannot be guaranteed. This can only be the case when the quality of crew schedules is explicitly considered while developing the solution method. (Steinzen, 2007), Ch. 6 is the only approach that does this. However, we notice that there are more authors who deal with irregular timetables (in total 5 out of 1820), but do not adapt their solution methods to take crew schedule regularity into account21. In other words, these approaches ((Borndörfer et al., 2004), (Kéri and Haase, 2007), (Bartodziej et al., 2009) and (Hollis, 2011), Ch. 4) will almost certainly produce irregular crew schedules, which are undesirable for – obviously – the crew but also for the company. Future papers should therefore pay more attention to the demand for more regular crew schedules when dealing with an irregular timetable. Although preventing delays from occurring should also certainly become a more investigated subject in the future, we see that yet more approaches are concerned with service level (3) than with quality of crew schedules (1). This is the same as saying that customers are generally considered to be more important than employees. On top of that, a higher service level is obtained by applying a dynamic approach, which implies that the schedules are only known just before they should be executed, asking a more flexible way of working from the staff. Those 20 In our categorization, a clear majority of the papers consider regular timetables. However, it should be noticed that since 2007, half of the approaches acknowledge the possible irregularity of timetables. We therefore assume that future publications will increasingly work with irregular timetables. 21 This usually includes solving the problem for a multiple period (e.g. a week) instead of a single period planning horizon (e.g. one day) (see section 1.1.1). One can imagine a transport company with different scenarios for weekdays and Sundays, where planning for a single period would mean scheduling weekdays and Sundays separately (in fact considering them as separate instances) so that the resulting schedules would be unrelated and the entire weekly schedule would not be regular. This pursued regularity is only possible when weekdays and Sundays are both considered in a multiple period planning horizon. 40 crew members might very easily become frustrated by such an approach. Naturally customer remains king, but it would not be a bad thing to tilt the attention a little more towards the employees wellbeing. The vast majority of VCSP papers apply the proposed approaches to real-life problems (15 of the 18 papers are application-based). This practically oriented view can only be encouraged and is certainly the path to be followed in the future as well. Furthermore, theoretical approaches should also be tested on real-world instances to prove their practical relevance and applicability. From our survey, it appears that the largest instance tackled so far is of the size of 1,414 trips. This is already quite large, although in practice there do exist transportation networks comprising several thousands or even tens of thousands of trips22. Up to now, these instances are not tractable as a whole, or at least no attempt has been made. It should be an objective to keep pushing the size boundaries so even huge real-world problems can maximally benefit from integrated approaches. So far, this seems not to have been a primary objective, since instances have not really increased in size from 2004 to 2011 (the largest instance already dates back to 2004, considered in (Borndörfer et al., 2004)). The last statement is true when we consider the number of trips as the absolute measure of size for a problem. However, we should also look at the number of depots and vehicle types considered, because these characteristics are just as well an aspect of the ‘size’ of an instance and actually increase the veracity of the problem (e.g. a transport company that uses only one or two vehicle types is not always realistic). In this respect, we do notice an evolution towards more extensive instances: (Bartodziej et al., 2009) consider no less than 27 vehicle types while (Hollis, 2011), Ch. 4 introduces a staggering number of 23 depots, also considering multiple vehicle types. Therefore, increasing the number of depots and vehicle types does not seem to be a bottleneck for expanding the tractable size of instances, the emphasis should be on the number of trips. It is noteworthy that we consider number of vehicle types as a measure of size and thus complexity of the problem, whereas number of crew types is not mentioned. This is because, in almost all cases, only one single crew type is assumed (so placing a ‘1’ in every column for number of crew types would not contribute a lot to the categorization). This is of course not very realistic, since employees can differ significantly with respect to skills, but also rights and responsibilities. For example, in an ex-urban distribution context, an older driver may have earned the right to drive the shorter routes so that he does not have to spend a night away from home, in contrast with junior employees. Seniority is only one criterion for defining different crew types, one can easily think of many others. When dealing with irregular timetables, a distinction 22 Our figures seem to demonstrate that a whole network of a respectable transportation company consists of at least a thousand trips. 41 between crew types can also be relevant for the objective of quality (regularity) of the crew schedule. For example, senior employees are often spared from irregular working hours, while starters should be willing to work overtime. Thus, a difference in importance of constructing a regular crew schedule should be implemented, depending on crew type. It should be clear from this exposition that the consideration of multiple crew types would be a meaningful addition for future papers. It is clear that most of the VCSP approaches are situated in an ex-urban context (12 out of 18, plus 2 approaches that consider both urban and ex-urban problems). The main reason for this observation can be found in (Huisman, 2004), stating “that it is generally expected that the savings of using an integrated approach (VCSP) in an ex-urban setting are much more significant than in an urban context, since there are much less opportunities to relief one driver for another one. These reliefs are only allowed at depots and certain other specified locations, which are much further away from each other than in the urban context. If first an optimal vehicle schedule is constructed (i.e. traditional sequential approach), there can be vehicles which do not pass a relief location for hours. Therefore, it is very well possible that there does not exist a feasible crew schedule at all, or more probably, that the crew schedule will be very inefficient.” (Huisman, 2004, p. 77) That is why an integrated approach is almost obligatory for ex-urban scenarios and should thus be thoroughly investigated (more thoroughly than urban scenarios). In more than two thirds of the cases (13 out of 18), changeovers are allowed. This can be explained by the fact that for larger problems (say of more than a couple of hundred trips) allowing changeovers can save many drivers and vehicles, whereas an integrated approach gives the same results for the cases with and without changeovers only for very small problems ((Huisman, 2004), Ch. 3.4). Of the scenarios where changeovers are allowed, only 2 consider unrestricted changeovers. It is a relatively underexposed characteristic to not impose restrictions to changeovers. Initial research showed that better and faster solutions can be obtained for instances with more than 80 trips if changeovers are not restricted ((Steinzen, 2007), Ch. 3.4). Still, we reckon further research is needed so we can be more conclusive regarding this aspect. It could be suspected beforehand that almost all post-2004 approaches make use of complete integration. In fact, only one approach still relies on partial integration ((Steinzen, 2007), Ch. 6). The reason (Steinzen, 2007), Ch. 6 uses a partially integrated method – somewhat outdated and less performant, but easier to implement than a fully integrated method we might state – could be found in the fact that it is the only approach that tries to construct regular crew 42 schedules from an irregular timetable, which is an innovative and non-straightforward objective. Since this specific problem focuses strongly on the wellbeing of the crew, it is not surprising that the crew is scheduled first and independently of the vehicles. In the future though, the problem should also be tackled using a completely integrated approach. Other future papers are evidently also expected to rely on a full integration of vehicle and crew scheduling. It was already mentioned that VCSP approaches which tackle the whole existing transportation network at once consist of at least a thousand trips in our categorization. We detect 3 of these approaches. We also identify 4 methods that solve each route separately and 6 that consider part of the network. Obviously, these categorizations strongly depend on the specific real-life problems that are considered in the paper (e.g. for a transportation company that operates about 1,000 trips it will be easier to tackle the whole network at once than for a company that operates 2,000 trips), but still they can give some insight of the scope of an approach. It should be the eventual goal of every transport company to solve their whole network at once. In that way, all possible elements (trips, depots, vehicle types, crew types, etc.) are taken into account all together so that a maximum effect of integration can be obtained. The set partitioning formulation is clearly the most popular one, it is used in 12 approaches, of which half are pure set partitioning. In other ‘non-pure’ cases, set partitioning is combined with another formulation type, mainly multicommodity flow. This multicommodity flow formulation is mostly used only for the vehicle scheduling part of the model (while set partitioning is then used for the crew scheduling part), which explains why only one approach employs a pure multicommodity flow. Set covering seems to be quite a lot less popular than set partitioning. Although, set covering formulations have the advantage that the number of variables, as the solution space, is considerably reduced. With a view to solvability, it could thus be advisable to base the model on a set covering formulation for larger and more complex future problems (high number of trips, depots, vehicle types, crew types). The largest instance tackled so far considering all measures of size combined (number of trips, depots and vehicle types) was introduced by (Hollis, 2011), Ch. 4 and is indeed modeled using set covering, supporting the previous argument. A more exotic formulation is the constraint-based one proposed by (Laurent & Hao, 2008), which is specifically designed to provide a flexible basis for the introduced metaheuristic (GRASP). We do not elaborate too extensively on the various algorithms used, as we did not explicitly present the mathematical formulas describing them fully (remember this was not within the scope of this thesis). What we can observe, is that column generation is clearly the most popular component of a solution approach. No less than 14 (out of 19, and not 18, because 43 (Bartodziej et al., 2009) propose two algorithms) solution methods are based on column generation while in another two approaches column generation – which is a mathematical programming method – is one of the two parts of a hybrid algorithm. This is not surprising, since it is so that to solve a (relaxation of a) set partitioning or set covering problem – which are used in almost every approach but a few – one has to resort to column generation (Huisman, 2004). Indeed, the two rather unusual models – the constraint-based by (Laurent & Hao, 2008) and the pure multicommodity flow by (Sato et al., 2009) – are not solved using column generation, but with a novel metaheuristic (GRASP) including constraint programming and a heuristic flow modification technique, respectively. Thus, non-mathematical programming methods are not so well established yet. We can distinguish between heuristics and metaheuristics23. We identify two methods that are more or less situated in the first category, namely the hybrid algorithm of (Rodrigues et al., 2006) comprising a greedy heuristic and the heuristic flow modification approach of (Sato et al., 2009). Also, three metaheuristics are proposed: a hybrid evolutionary algorithm by (Steinzen, 2007), Ch. 4, a so-called GRASP by (Laurent & Hao, 2008) and some local search-based metaheuristics (SA, GDA, RRT) by (Bartodziej et al., 2009). The hybrid evolutionary algorithm of (Steinzen, 2007), Ch. 4 was actually the first attempt to apply a metaheuristic to the VCSP. Since then, some other metaheuristic approaches have been proposed, but still it remains an area for which further research is necessary. We should seek to answer questions such as ‘Which particular metaheuristic performs best (for which particular problem)? 24’ and ‘Can we make use of metaheuristics to solve even larger/more complex reallife problems (compared to mathematical programming25), and if so, how much larger/more complex may those problems be?26’ in order to discover the true potential of metaheuristics. Until now, the practical problems tackled with a metaheuristic (or heuristic or hybrid) approach 23 For a comprehensive overview of solution methods (comprising constraint programming and mathematical programming approaches) we refer to (Ernst et al., 2004). A clear distinction between the nature of (classical) heuristics and metaheuristics can be found in (Toth & Vigo, 2002). These authors also state that metaheuristics outperform classical methods in terms of solution quality (and sometimes even in terms of computing time), so there is little room left for significant improvement in the area of classical heuristics and thus future focus should be on metaheuristics. 24 E.g. according to (Toth & Vigo, 2002) (they consider the VRP, not the VCSP), tabu search emerges as the most effective approach. Procedures based on pure genetic algorithms (a subclass of evolutionary algorithms) and on neural networks were clearly outperformed, while those based on simulated or deterministic annealing and on ant systems were not quite competitive at the time (remember this was about 10 years ago). Although it is stated that hybrid ant systems and genetic algorithms might, in the future, be able to match the effectiveness of existing tabu search heuristics. Also, (Bartodziej et al., 2009) conclude that the metaheuristics considered by them, “SA, GDA and RRT, show a relatively similar convergence. […] Although, RRT has the fastest convergence at the beginning and is only caught up if the running times are relatively large. […] On the other hand, SA and GDA are slightly preferable when the first objective is to minimize the number of auxiliary vehicles.” (Bartodziej et al., 2009, pp. 426-427) Finally, preference was given to RRT. 25 Generally, algorithms based on a mathematical programming approach will still achieve the lowest cost solutions. 26 (Bartodziej et al., 2009) give a first indication of the truthfulness of this statement, since the largest instances (> 695 trips and up to 1,400 trips) considered could not be solved by column generation, but were indeed solvable using metaheuristics. 44 are overall not more complex (not only regarding the known measures of size, but also considering e.g. the admission of changeovers) than the ones that are solved using column generation. In fact, rather the opposite is true. Also, the capabilities of constraint programming to solve highly constrained, thus very complex, transportation problems should be investigated. A remarkable aspect concerning the algorithms based on column generation is that, although linear relaxation was said to be more popular (see section 1.1.3), up to 2007 almost all presented VCSP approaches used a Lagrangian relaxation to solve the master problem. This could be explained by the fact that (Huisman, 2004) chose to relax the proposed formulation in a Lagrangian way. Being a real landmark paper, several authors who published in the next few years based their approach on that of (Huisman, 2004) (e.g. (Steinzen, 2007)). Since 2008, however, LP relaxation is again most widely used, reclaiming its popularity. Regarding the construction of feasible solutions (again for algorithms using column generation), Lagrangian heuristics and branch-and-bound methods appear to be the preferred approaches. Furthermore, the latter seem to have become the most popular nowadays. This evolution can be explained by the fact that a higher quality of the algorithm (i.e. closing the gap with the lower bound) can be obtained by using an exact (e.g. branch-and-price) method instead of Lagrangian heuristics. Some other feasible solution construction methods have also been proposed, as there is a heuristic branch-and-price method in (Steinzen, 2007), Ch. 3.1-3.3, a method based on local and follow-on branching in (Steinzen, 2007), Ch. 6 and a round-up method in (Kéri & Haase, 2007). We already mentioned that the VCSP is solved dynamically in only 2 of the 18 approaches. Now what is the reason that this topic has not yet received that much attention? It was proved that for small instances with a single depot, a dynamic approach performs well to reduce delays. However, computation times were still high for applying such an approach in practice. On the other hand, for medium-sized instances with multiple depots – which is a much more realistic situation – a traditional static approach with inclusion of buffer times performed much better. ((Huisman, 2004), Ch. 5.3) This might indicate that some authors presume that the idea of dynamically solving itself does not work so well. Nevertheless, it is recommended to invest further research in speeding up the suggested algorithms. With faster computers and better algorithms the dynamic approach could outperform the static one with buffer times for larger problem instances as well. However, to the best of our knowledge, no existing paper has yet explicitly showed that a dynamic approach may also be beneficial for larger instances, compared to a traditional static approach with buffer times. Previously, we related the dynamic approach to the service level objective which aims at reducing delays for customers. This is certainly true in public transport, but does not completely hold in a distribution context. For a distribution company, in fact, the prevention of delays is far 45 less pertinent because of the existence of time windows for deliveries, so that a certain delay tolerance is already present. A dynamic approach in the product delivery context could instead be understood as the possibility of including a new customer in a predefined route, during the same day that particular customer places his order. A situation like this occurs in (Bartodziej et al., 2009): “during operation of the fixed timetable, airlines may ask the trucking company for additional transportation tasks on the spot.” (Bartodziej et al., 2009, p. 406) In such situations, a dynamic scheduling approach would indeed be useful in order to rearrange the existing schedule in such a way that it remains not far from optimal under the changed circumstances. Such a dynamic approach in the distribution context we did however not yet encounter, but it could be an interesting addition to the existing literature. It is not surprising that virtually all authors present a general conclusion stating that their approach can get good solutions (mostly also for specific real-world instances of considerable size) within reasonable computation times. In fact, this is and should always be an obvious goal when designing a new solution method. More importantly, we now have enough information to solve the central research question, or at least the first aspect of it, which regards the scheduling problem27: ‘Is it (always) advantageous to use an integrated approach for scheduling problems in transportation?’ Actually, we should not make too much fuss about how to respond to this question, because the answer is clearly and indisputably: ‘Yes’. We can give many literal quotes from the presented VCSP approaches that evidence – and none that contradict – the positive answer, as we will do here: ï‚· “The main conclusion is that we can save vehicles and/or crews by integrating the vehicle and crew scheduling problem, which may lead to a big decrease in costs.” (Huisman, 2004, p. 76), Ch. 3.3 and Ch. 3.4 ï‚· “There are significant savings compared to the traditional sequential approach, where first the vehicle scheduling and afterwards the crew scheduling problem is solved.” (Huisman, 2004, p. 100), Ch. 4.2-4.3 and Ch. 4.4 ï‚· “The solutions produced can be decidedly better in several respects at once than the results of various types of sequential planning.” (Borndörfer et al., 2004, p. 20) 27 Recall that the second aspect of the central research question concerns the combined routing and scheduling problem and will be considered in the next section. 46 ï‚· “There is an efficiency gain compared to sequential planning. […] Even the number of vehicles is always minimal, i.e., equals the number of vehicles when sequential planning is performed, where vehicles are scheduled first (and therefore are optimal).” (Steinzen, 2007, pp. 116-117), Ch. 3.1-3.3 ï‚· “The results show that there is an efficiency gain if vehicle and crew scheduling are treated in an integrated way.” (Steinzen, 2007, p. 121), Ch. 3.4 ï‚· “The approach discloses significant savings compared to the traditional sequential approach without requiring a fully integrated solution method.” (Steinzen, 2007, p. 135), Ch. 4 ï‚· “First, one observes that the integrated approach always outperforms the sequential one, or at least furnishes equivalent results. In particular, the savings in terms of number of drivers are significant […]. The sequential approach provides a lower bound for the number of vehicles that is always reached in the integrated solutions. […] Second, the integrated approach is more powerful than the sequential one in the sense that the sequential approach failed to solve a particular instance where the integrated approach succeeded. […] These results show the dominance of the integrated approach over the sequential one.” (Laurent & Hao, 2008, p. 474) Notice that, except for (Laurent & Hao, 2008), none of the most recent VCSP papers (say dating from the last 4 to 5 years) directly compare their approach with a traditional sequential one, although they do make a comparison with other existing integrated approaches (see below). This could be perceived as the existence of the general conception that, whatever the specific problem is, an integrated VCSP approach will always prove to be more advantageous than a sequential one, so that an explicit comparison is not useful anymore. Moreover, it is self-evident that when a particular integrated method is more efficient than another earlier one – which has been proven to outperform the sequential approach – that new method will of course also be more beneficial than the sequential approach. Given the amount of VCSP approaches we considered in this thesis – we did not just evaluate the performance of one single proposed method relative to the traditional sequential approach – we may assume to have the ‘right’ and the opportunity to state that it is always beneficial to use an integrated approach for transportation scheduling problem, which can be seen as an innovative contribution of this thesis. 47 Although the use of an integrated method is always recommended, it can prove more or less beneficial depending on the specific conditions, reflected by the different criteria. For example, we already mentioned in this section that it is generally expected that the savings of using an integrated approach (VCSP) in an ex-urban setting are much more significant than in an urban context. (Laurent & Hao, 2008) confirm this by stating that “the integrated approach is indispensable especially when relief opportunities are rare” (Laurent & Hao, 2008, p. 475), so in an ex-urban context. There are still other conditions that may influence the degree of profitability of an integrated approach. We will list some of them here. ï‚· “The interpretation of the results depends on the ratio between the fixed vehicle and crew costs. If fixed vehicle costs are much higher as compared to crew costs it becomes less attractive to apply the integrated approach. On the other hand, if crew costs are higher the integrated approach becomes more attractive.” (Huisman, 2004, p. 62), Ch. 3.3 ï‚· “In the case that no changeovers are allowed, the benefit of integration may be very significant because more vehicles and/or crews can then be saved.” (Huisman, 2004, p. 76), Ch. 3.4 ï‚· “When we do allow changeovers, it is possible to reduce the total costs by allowing changeovers more often, so by making the changeovers less restricted.” (Huisman, 2004, p. 76), Ch. 3.3 (Steinzen, 2007), Ch. 3.4 confirms this by stating that “it is worthwhile for planners in practice to allow unrestricted changeovers since the additional flexibility results in efficiency gains.” (Steinzen, 2007, p. 121) The reader must not forget that an integrated approach is still advantageous over a sequential one, even if the conditions are such that only a lower profitability can be obtained by integration. Some authors do in fact explicitly compare their approaches to other existing methods. We will not present these results here (interested readers are referred to the conclusions of the comprehensive categorization in Appendix E), because it falls outside the scope of this thesis to identify the ‘best’ available method at this time. It should be mentioned that – besides the fact that not all authors directly compare their approach to others – another important difficulty in the search for the ‘best’ VCSP method is the absence of lower bounds in more than a few approaches. For instance, (Rodrigues et al., 2006) do not compute lower bounds and, consequently, the quality of their solutions cannot be 48 assessed. Also for (Kéri & Haase, 2007) it should be noticed that the approach was not compared with any other existing method from literature, nor were lower bounds calculated. Just as for (Laurent & Hao, 2008), who themselves mention that “a more complete assessment would compare the results with tight lower bounds, which are unfortunately unavailable yet.” (Laurent & Hao, 2008, p. 474) Therefore, identifying the ‘best’ integrated scheduling method for transportation problems is left to future papers. 2.2. Evolution of the VCRSP A Vehicle and Crew Routing and Scheduling Problem (VCRSP; see section 1.4) concerns the combination of the routing problem and the scheduling problem, which only occurs in the distribution context (e.g. freight distribution, postal organizations and delivery services). We do not encounter VCRSP’s in a public transport context, because line routes (and generally also timetables28) are fixed here, dictated for example by a local authority. In section 1.4, the subject of VCRSP was not extensively discussed because no relevant pre-2004 papers were to be found. The last few years though, some papers have been devoted to the planning process in a distribution company. We will discuss these papers chronologically in order to give an overview of the evolution in this area of research. More specifically, we examine to what extent and in which way the routing aspect and the scheduling aspect are integrated. Recall that we are now dealing with an integration that is higher-level than the one in the previous section, which was limited to the integration within the scheduling problem (or in other words, the VCSP). Note that the methods presented here will not be categorized in the same extensive manner as we did for the VCSP methods in section 2.1, mainly because only few papers on the VCRSP exist so that an extended categorization would not (yet) be very significant. Still, we can provide a meaningful two-dimensional classification for the discussed VCRSP approaches. This section is divided into following subsections. First (section 2.2.1) we provide an overview of existing approaches for the planning problem in the distribution context (including a twodimensional classification) and afterwards (section 2.2.2) we draw conclusions based on this overview, identifying gaps within the research area with the aim of detecting topics for future research. Finally, we try to answer the central research question related to the VCRSP. 28 This was not the case in (Rodriguez et al., 2006) and (Kéri & Haase, 2007) discussed in the preceding section 2.1, where the timetable was prepared during these approaches. This comes down to the inclusion of the timetabling phase (which is sometimes seen as a part of the operational phase – see section 1.2 – and therefore its occurrence in a scheduling problem is partially explained) in the solution approach. Of course, once the timetables are defined (timetables apply for a long period in public transport), we fall back to a standard VCSP. 49 2.2.1. Overview of existing approaches In order to provide a starting base, we repeat the definition of the traditional sequential approach for routing and scheduling. The traditional planning process for distribution problems typically involves constructing a set of vehicle routes, then producing vehicle schedules by assigning vehicles to the vehicle routes, and building a set of crew schedules so that all vehicle routes are covered. (Hollis, 2011) As was also indicated by (Drexl et al., 2011), most recent solution methods for distribution problems are based on the recurring idea of decomposing the considered problem into several stages and, in first stages, take some aspects into account which are needed for obtaining feasible solutions in later stages (which actually comes down to partial integration). Yet, different approaches in doing this are possible. As already mentioned in the introduction of this section, we can identify two dimensions to categorize the different approaches. (Drexl et al., 2011) provided us with the inspiration for this classification, but did not develop an actual categorization themselves, nor did they make the connections we will present here. A first dimension for categorization concerns the allowance of changeovers. On the one hand, there are papers in which drivers can change vehicles only at a central depot. Obviously, this comes down to a single depot case, more specific in an ex-urban context (where distances are such that changeovers are often restricted to only take place at the depot(s)). On the other hand, there are papers where changeovers are allowed in other relief points as well, possibly but not necessarily depots. Normally, this comes down to a multiple depot case, although an arrangement with a single central depot supplemented by several other relief points – that are not depots – is also possible. The locational exibility of the second class adds an additional degree of freedom. Hence, problems in the second class turn out to be significantly harder to solve than those in the first class. A second dimension regards the solution method used. Three approaches can be identified: (1) first determining routes and assigning concrete vehicles and crew members afterwards, (2) calculating routes for predetermined crew-vehicle pairs, and (3) performing a direct selection of vehicles, crews and tasks. Remark that (2) actually takes the opposite way of (1), with respect to the sequence of routing and scheduling. We point to the clear similarities of this classification with the ‘degree of integration’-criterion for the VCSP approaches introduced in section 2.1.2.1. (1) and (2) could both be seen as partial integration approaches, while (3) is more of a strive for complete integration. When we make a further subdivision, the routing first - scheduling second 50 approaches of (1) are parallel to vehicle first - crew second for a VCSP, and the scheduling first - routing second of (2) to crew first - vehicle second. Notice that the two defined dimensions coincide perfectly with the two overarching categorization aspects – practical problem and solution method – defined for the VCSP in section 2.1. The first dimension deals with the admission of changeovers, while the number of depots and degree of urbanization are also mentioned, all known criteria for the practical problem. It was already explicitly mentioned that the second dimension regards to the solution method used, more specifically the degree of integration. These observations prove that it would certainly be possible and quite straightforward (due to the similarities) to design a categorization for the VCRSP that is parallel and just as elaborate as the one introduced for the VCSP in previous section. As already said, the existing literature on the VCRSP might not yet be extensive enough to develop such an elaborate categorization at this time, but it certainly is an interesting possibility for the future. The relevant recent publications on distribution problems that we will consider here are, in chronological order: (Hollis et al., 2006), (Xiang et al., 2006), (Laurent & Hao, 2007), (Zäpfel & Bögl, 2008), (Kim et al., 2010), (Prescott-Gagnon et al., 2010) and (Drexl et al., 2011). This selection is not arbitrarily, but corresponds to the VCRSP publications deemed most important by the last mentioned authors. In Table 2, we categorize the papers according to the two defined dimensions. partial integration changeovers only at central depot changeovers at multiple relief points complete integration routing - scheduling scheduling - routing (Xiang et al., 2006) (Zäpfel & Bögl, 2008) (Laurent & Hao, 2007) (Prescott-Gagnon et al., 2010) - (Hollis et al., 2006) (Drexl et al., 2011) - (Kim et al., 2010) Table 2: Two-dimensional classification of relevant VCRSP papers A brief chronological overview of above papers, describing the specific problems that are considered and the way these problems are tackled, is included in Appendix F. Since (Drexl et al., 2011) already presented such an overview, it would not be a contribution to do the exact same thing ourselves. For the sake of completeness and to make verification of our presented two-dimensional classification possible, we chose to include the (Drexl et al., 2011) overview in Appendix F nonetheless. Also, the results from computational studies mentioned in the 51 concerning papers themselves were added. We will need those to draw our conclusions in next section, however the relevant parts will also be cited there. 2.2.2. Conclusions It is noteworthy that, although it are always vehicles and drivers that need to be scheduled in the papers, the concrete application contexts are almost all different (limousine rental, mail distribution, oil delivery, just to name a few). Consequently, we could say that a fairly wide research area is already being covered, regarding the considered practical problems. It is certainly a good thing that this kind of variation is propagated by the different papers, so possible benefits of integration can be obtained for every particular distribution problem. On the other hand, VCRSP methods that are initially designed for one specific application (e.g. oil delivery) and perform well for it, should be adapted in order to make it more generally applicable for other practical problems (e.g. mail distribution) as well. This defines a first recommendation for further research. Although the categorization of the papers in Table 2 is rather limited, some conclusions can be derived from it. First of all, we see that most papers consider a setup where changeovers are only allowed at the central depot. This single depot case is mostly assumed because it is remarkably easier to solve than the scenario with admittance of changeovers at multiple relief points, usually depots. Multiple depot instances are however very common in a distribution context and are furthermore more comprehensive. That is why it is recommended for upcoming papers to focus somewhat more on multiple depot problems, where changeovers are not constricted to only take place at a single location. When we take a look at the other dimension, it immediately stands out that virtually every publication on the VCRSP so far uses a partially integrated approach. The routing first scheduling second approach proves to be the most popular one. This is not surprising, since it corresponds best to the natural flow of the traditional planning process for distribution problems. A gap is identified for the scheduling first - routing second methods that solve a problem where changeovers are allowed at multiple relief points. This void should be filled, evaluating whether such an approach is more or less beneficial than one based on routing first - scheduling second. An advantage might indeed show, given that crew costs are mostly dominant (certainly for road traffic) so that scheduling of crews becomes the most important aspect, meaning it would be beneficial to construct the crew schedules as early on as possible in order to attain a better approximation of optimality. However, a more important gap to be filled is that of VCRSP 52 approaches that pursue a complete manner of integration. Up to now, only one paper has made such an attempt, namely (Kim et al., 2010). Although immediately considering the more difficult case where changeovers are allowed at multiple relief points, the authors do not present any relevant computational results and moreover do not develop any good lower bounds for the problem. Consequently, the solution quality of the proposed algorithm cannot yet be evaluated. Testing the approach of (Kim et al., 2010) on a practical instance is an idea for future research, but more importantly, other approaches pursuing complete integration should be developed (and applied, followed by an evaluation). It does not take a lot of insight to detect that the approaches for the VCRSP clearly lag behind those for VCSP, not only in number but also and foremost in terms of completeness of integration. Whereas almost all recent VCSP papers present an approach based on complete integration, for VCRSP papers this is actually a virtually non-existent research area. Of course, distribution problems are more extensive since the routing of vehicles has to be performed as well, on top of the scheduling of those vehicles and the crews. In section 1.2 we also mentioned that public transport problems are generally seen as more important than distribution problems because of the presence of certain specific difficulties (e.g. a more delicate trade-off between customer and cost), which too could explain and to some extent ratify the lead of VCSP approaches on VCRSP methods. Nonetheless, we feel it is time to dedicate some extra attention and effort to the VCRSP, since distribution problems are all the same an import aspect of today’s transport issues. (Hollis et al., 2006) actually provide a perfect indication of the current subordination of integration between routing and scheduling (VCRSP), as they consider a distribution problem, but with a focus on the introduction of a new VCSP method, only useful for the scheduling part of course. Although we managed to identify some topics for further research, the categorization in Table 2 is to limited to be able to recognize all the interesting future possibilities. An extension of this categorization towards a one as presented for the VCSP approaches in section 2.1 was already said to be possible, and would indeed prove very useful to identify more and clearer gaps in the research area and therefore provide a base for more comprehensive recommendations for future investigations. The design of such an extended categorization for VCRSP’s can thus in itself be understood as a subject for future research. To conclude, we try to answer the second aspect of the central research question29, concerning the combined routing and scheduling problem: ‘Is it (always) advantageous to use an integrated approach for combined routing and scheduling problems in transportation?’ 29 The first aspect of the central research question concerned just the scheduling problem (VCSP) and was already answered in section 2.1.4. 53 First of all, let us remark that providing a conclusive answer to this question is not so easy as for the VCSP, merely because of the fact that far less papers are available on the VCRSP. Another significant problem is that some authors do not even give any relevant prove of their VCRSP approaches merely being ‘good’. For example, (Kim et al., 2010) present no computational results whatsoever and admit that good lower bounds for the problem still need to be developed to evaluate the solution quality of the proposed algorithm. They only show that their proposed approach outperforms a simple greedy algorithm, which does not say much about the actual solution quality. Also (Prescott-Gagnon et al., 2010), who do obtain computational results on some instances derived from a real dataset, only mutually compare their own proposed methods, without testing them to existing approaches or providing lower bounds. Again, this does not really prove anything about the actual solution quality. Most of the authors though, do more or less evidence that their approach is ‘good’. Some literal quotes: ï‚· “The solution technique employed […] has been shown to find high quality solutions. […] Moreover, Australia Post is currently using software developed based upon the ideas presented in this paper to assist in the management of ongoing changes to the mail distribution networks in major cities throughout Australia.” (Hollis et al., 2006, p. 149) ï‚· “The performance of the heuristic was evaluated by intensive computational tests on some randomly generated instances, revealing that this method is capable of quickly obtaining acceptable results. Moreover, the method is flexible to cope with many practical constraints and does not contain any case-sensitive empirical parameter.” (Xiang et al., 2006, pp. 1135-1136) ï‚· “Within a short time, the software supplies very good quality schedules in which the major part of the trips is assigned, satisfying all constraints. […] The approach also proves to be flexible. […] Moreover, the decision support system based on this research is operating in the examined company and proves to be extremely useful.” (Laurent & Hao, 2007, p. 557) ï‚· “Computer simulations have demonstrated that the constructed heuristic is suitable for practice.” (Zäpfel & Bögl, 2008, p. 980) “All in all, the tabu search procedure can solve a complex decision problem with a tremendous number of variables and constraints in an efficient manner.” (Zäpfel & Bögl, 2008, p. 995) 54 ï‚· “First and foremost, the presented algorithm is shown to be capable of solving large realworld instances and is able to achieve consistent and practically relevant results.” (Drexl et al., 2011, p. 16) Of course, the central research question does not ask whether an integrated approach is ‘good’, but whether it is better than a traditional approach. Only two papers respond to this question, in some sense, explicitly: ï‚· “The algorithms and solution techniques presented in the paper have been used by network planners at Australia Post to demonstrate a potential transport network operational cost saving of 10% compared to the actual practice for the 2003 Melbourne metropolitan mail distribution network.” (Hollis et al., 2006, p. 149) ï‚· “Results obtained on real data sets show a significant improvement in terms of quality, operational costs and elaboration time, compared with the actual practice in the examined company.” (Laurent & Hao, 2007, p. 557) The actual practice for both Australia Post (Hollis et al., 2006) and the examined company in (Laurent & Hao, 2007) relied on the traditional planning process for distribution problems described at the very beginning of section 2.2.1. So indeed, we can state that these two papers prove that an integrated approach for the VCRSP is more advantageous than a traditional one. In other words, it looks as if an integrated approach for routing and scheduling has indeed the potential to be beneficial, but for the present, too little proof is provided to make this into a general conclusion. Moreover, there is still the aspect of the integrated approach always being advantageous. (Drexl et al., 2011) actually come to the conclusion that, for their given data set, lorry/driver changes offer no savings potential, so that making a fixed lorry-driver assignment would be the best choice. This in fact means that, for their specific problem, solving it as a standard VRP is more beneficial than approaching it as a VCRSP. However, the authors stress that these results are only valid for the considered business field and that the developed algorithm may well lead to very different results with other data or in other application areas. Anyhow, the statement of an integrated approach for combined routing and scheduling always being more advantageous seems to have been invalidated. Summarizing, there certainly is potential in the integration between routing and scheduling, at first sight less pronounced than for the VCSP, but in any case more research in the area of VCRSP is needed in order to evolve towards a universal conclusion. This further research can thereby lead to new methods (those existing today are in fact only the first steps) which may 55 well ever produce better results than the traditional approach, so we could then wholeheartedly answer ‘Yes’ to the central research question, just as for the VCSP. 56 3. Future research As already mentioned, it is beyond the scope of this thesis to propose improvements for specific existing solution methods, more particular for models and algorithms. We rather identify more general and problem-inherent future research topics. Those were already deduced and presented in sections 2.1.4 and 2.2.2 for the VCSP and the VCRSP, respectively. We will now repeat the resulting recommendations in an orderly fashion (in order of increasing specificity, so starting with the most general topic) and also mention – when available – some quotes for each topic, literally extracted from the considered papers (on VCSP as well as on VCRSP) that substantiate the relevance of the suggested recommendations for future research. First for the VCSP: ï‚· Continue on the path of mainly application-based papers and test initially theoretical approaches on real-world instances to prove their practical relevance and applicability. - “There can be quite some practical situations, which require more research.” (Huisman, 2004, p. 149) - “The intention now is to submit both the presented algorithm and the interface to intensive field tests at selected companies.” (Rodrigues et al., 2006, p. 861) ï‚· Future papers should all rely on a full integration of vehicle and crew scheduling. - “Computational results for the partially integrated (ex-urban) vehicle and crew scheduling indicate that the regularity can be improved while maintaining cost optimality30. However, we left a fully integrated consideration for future research.” (Steinzen, 2007, p. 179) ï‚· It should be the eventual goal of every transport company to solve their whole network at once so that a maximum effect of integration can be obtained. 30 This concerns the (Steinzen, 2007), Ch. 6 VCSP approach which was categorized in section 2.1.3. Indeed, this approach had the objective of producing regular crew schedules and was the only approach considered that relied on a partial integration. 57 ï‚· Shift the attention a little from customer satisfaction (service level objective) towards wellbeing of employees (quality of crew schedules objective), although customer remains king. ï‚· Adapt the presented categorization procedure of section 2.1 for airline problems and expand it in terms of number of considered papers for the other modes of transport, especially railways. ï‚· Focus more on ex-urban (than on urban) scenarios, since an integrated approach is almost obligatory in that setting. ï‚· Take into account the demand for more regular crew schedules when dealing with an irregular timetable. - “We deem it worthwhile to include timetable considerations (i.e. irregular timetables still leading to regular crew schedules) into the integrated treatment of multiple-depot vehicle and crew scheduling.” (Steinzen, 2007, p. 180) - “We suggest to continue research on aspects related to the quality of work conditions (i.e. regular crew schedules).” (Steinzen, 2007, p. 180) ï‚· Introduce a dynamic scheduling approach for the distribution context. - “A next step will be the integration of the system with the dispatching system, i.e. to allow feasibility checking and the proposal of optimal insertions of ad-hoc trips into RFS-plans31 in a dynamic environment.” (Bartodziej et al., 2009, pp. 428429) ï‚· Extend solution approaches designed for a single depot case to a multiple depot case, especially when a single depot was assumed to simplify the introduction of a new solution method. ï‚· Pay more attention to the development of dynamic scheduling approaches (to prevent delays from occurring), including explicitly showing that a dynamic approach may also 31 Remember a RFS is indeed a distribution service. 58 be beneficial for larger instances, compared to a traditional static approach with buffer times. - “We suggest to continue research on faster algorithms to solve the dynamic integrated vehicle and crew scheduling problem, since the dynamic approach was only able to solve small problem instances due to the large computation time. […]Moreover, we suggest to test such a dynamic approach in a real-world environment such that our assumptions can be checked and to see if such an approach also works in practice.” (Huisman, 2004, p. 149) - “Introducing the ideas of dynamic approach in airline and railway environments would be an interesting subject for future research.” (Huisman, 2004, p. 149) - “We suggest to continue research on aspects related to the quality of vehicle and crew schedules such as robustness32.” (Steinzen, 2007, p. 180) - “Let us mention that the general solution approach shown in the paper is also suitable for dynamic adjustment of schedules by local re-optimization.” (Laurent & Hao, 2008, p. 475) - “For future work, the proposed solution method needs improving of the search efficiency to handle larger disruptions.” (Sato et al., 2009, p. 150) ï‚· Keep pushing the size boundaries (especially number of trips, since increasing the number of depots and vehicle types does not seem to be a bottleneck) of instances so even huge real-world problems can maximally benefit from integrated approaches. - “Although some progress has been made over the past years, we are not aware of an approach that could deal with several thousands or even tens of thousands of trips. However, problem instances of such size with many depots are common in big cities such as the German towns of Munich, Hamburg, or Berlin. Therefore, we suggest to pursue further research on this topic.” (Steinzen, 2007, p. 180) ï‚· Consider multiple crew types instead of by default assuming only a single crew type. 32 A more robust solution is understood to be one where disruptions in the schedule (due to delays) are less likely to be propagated into the future, causing delays of subsequent trips. This exactly corresponds to the objective of dynamic scheduling. 59 - “Our approach can easily be extended to the case with non-identical crews.” (Steinzen, 2007, p. 62) ï‚· Further research is needed to be more conclusive regarding the possible advantage of allowing unrestricted changeovers (compared to restricted changeovers). - “For future work, there are several conditions to be considered according to each situation, which have not been covered by the proposed solution method. These include rides that crews take without driving or conducting (corresponds to making changeover less restricted).” (Sato et al., 2009, p. 150) - “An even more involved extension of our problem is to allow that a change of driver/lorry at a relay station is performed without the driver taking a daily rest before switching to another lorry (i.e. allowing unrestricted changeovers).” (Drexl et al., 2011, p. 18)33 ï‚· Identify the ‘best’ integrated scheduling method for transportation problems, which asks for the computation of (tight) lower bounds and/or the comparison of the solutions obtained by different approaches34. - “A more complete assessment would compare the results with tight lower bounds, which are unfortunately unavailable yet.” (Laurent & Hao, 2008, p. 474) - “It is also important to design a mathematical solution method based on the relaxation techniques. Because the mathematical method provides such a good lower bound of the optimal solution, we are able to numerically evaluate the capabilities of the proposed method.” (Sato et al., 2009, p. 150) ï‚· Base models on a set covering formulation to increase solvability of larger and more complex future problems. 33 This paper actually considers a VCRSP, but the specific aspect of changeovers is clearly mainly – if not only – related to the scheduling part of the problem, and thus to a VCSP. 34 Preferably calculated on the same computer, considering the exact same problem (or with only very little variations, possibly needed to make a specific algorithm applicable for the problem). 60 ï‚· Explore the capabilities of constraint programming to solve highly constrained, thus very complex, transportation problems. ï‚· Further research in the area of metaheuristics35 is necessary, not only to identify which particular metaheuristic performs best, but primarily to discover their true potential for solving larger/more complex real-life problems (compared to mathematical programming). - “Further research in the field of metaheuristics will focus on how to partition the trips assigned to a depot among vehicles and drivers with a local search heuristic.” (Steinzen, 2007, p.135) - “We can evaluate whether a mathematical method has the potential to be more suitable for the crew/vehicle rescheduling than the proposed heuristic method.”(Sato et al., 2009, p. 150) - “Further research directions include the application of other metaheuristics.” (Zäpfel & Bögl, 2008, p. 995)36 And for the VCRSP: ï‚· Dedicate some extra attention and effort to the whole of the VCRSP, since it is underdeveloped compared to the VCSP while nonetheless, distribution problems are an import aspect of today’s transport issues. ï‚· More general research on the VCRSP is needed for progressing towards a valid conclusion whether or not an integrated approach is (always) more advantageous than a traditional one. Authors should therefore explicitly compare their VCRSP method to the traditional approach (or to other VCRSP methods). ï‚· Develop, apply and evaluate approaches pursuing complete integration. 35 For completeness, we could in fact also consider (classic) heuristics. Although the authors actually consider a VCRSP, the metaheuristics are used for the scheduling part so the mentioning of the paper is relevant here. 36 61 - “Good lower bounds need to be developed to evaluate the solution quality of the proposed algorithm37 and future algorithms, which will be devised.” (Kim et al., 2010, p. 8430) ï‚· Test existing theoretical approaches on practical instances. ï‚· Design an extended categorization for VCRSP’s (Table 2 can provide a basis), analogous to that for VCSP’s of section 2.1. ï‚· VCRSP methods that are initially designed for one specific application and perform well for it, should be adapted in order to make it more generally applicable for other practical problems as well. - “Although only the static dial-a-ride problem is solved in this paper, some proposed techniques can be used to solve other problems.” (Xiang et al., 2006, p. 1136) - “This paper deals with a real-world driver and vehicle scheduling problem in a particular application context. Even if some aspects are specific, others are general ones. In particular, the notion of simultaneous scheduling of drivers and vehicles is quite general and relevant to many other scheduling applications.” (Laurent & Hao, 2007, p. 557) - “The current problem can also be extended in many ways. These include introducing time window constraints requiring that the service at each customer starts within an associated time window.” (Kim et al., 2010, p. 8430) - “As a future research direction, one can consider extending the proposed heuristics to treat a multiple product version of the problem where the vehicles have several compartments of fixed sizes.” (Prescott-Gagnon et al., 2010, p. 14) - “The developed algorithm for simultaneous vehicle and crew routing and scheduling may well lead to very different results with other data or in other application areas, so that its further study is justified.” (Drexl et al., 2011, p. 18) 37 The algorithm proposed by (Kim et al., 2010) does indeed rely on complete integration. 62 ï‚· Focus more on multiple depot problems, where changeovers are not constricted to only take place at a single location. - “Further research directions include the extension of the solution concept to multi-depot combined vehicle routing and personnel planning problems, which arise when connected regions of a logistic service provider must be included in this context.” (Zäpfel & Bögl, 2008, p. 995) - “A deeper study of problems where elementary objects may join and separate on the fly at many different locations constitutes a challenging research area.” (Drexl et al., 2011, p. 18) ï‚· Fill the void of scheduling first - routing second methods for solving a problem where changeovers are allowed at multiple relief points, evaluating whether such an approach is more or less beneficial than one based on routing first, scheduling second. We already indicated that and explained why the VCRSP recommendations for future research are quite a lot less comprehensive and concrete than those for VCSP. Now this also clearly shows from the above listings. By adequately following the first and fifth recommendation for the VCRSP, this gap can however quite easily be closed. Indeed, more relevant papers on the VCRSP will ensure a broader base for coming to the right conclusions and an extensive categorization of these papers may lead to the identification of many concrete research topics, as was the case for the presented VCSP categorization. Notice that almost half of the suggested research topics were not explicitly mentioned in the considered papers. This certainly does not mean that these recommendations are irrelevant. In fact, there are three equally important reasons which ratify their absence. A first one is that we also included recommendations that are most probably already known and followed by the collection of authors and were thus not explicitly mentioned by them (e.g. focus more on exurban (than on urban) scenarios, since an integrated approach is almost obligatory in that setting). We however did indicate these topics for the sake of completeness. Secondly, some subjects may be ‘new’ because we used a very broad and general view on the planning process in transportation, whereas the vast majority of papers focuses on a rather specific problem. For instance, research issues such as shifting the attention a little from customer satisfaction towards wellbeing of employees can only be identified when looking at the bigger picture. And third, a few topics were only revealed through the introduction of our own categorization 63 procedure, so will of course be presented for the very first time in this thesis. An example is the recommendation for filling the void of scheduling first - routing second VCRSP methods for solving a problem where changeovers are allowed at multiple relief points. Furthermore, it is remarkable that many authors seem to acknowledge the possibility and importance of dynamically approaching the scheduling problem. However, we observed that only very few papers actually present such a dynamic approach. We have thus identified a high need and a low presence, so consequently, dynamic scheduling should be treated as one of the hottest amongst the topics. For the VCRSP, the generalization of methods specifically designed for a particular problem, so that they can be applied for other problems as well, is the most frequently mentioned research subject. This is of course not surprising, since it concerns a very general recommendation that actually reflects the logical lifecycle of a newly developed solution method. We have to remark that not all research topics will be completely satisfiable at this point, because quite a few of them entail an increase in problem complexity and thus a decrease in tractability. These difficulties can only be overcome by faster computers and better algorithms. The latter aspect is an obvious point of research (e.g. “the quality of the algorithms to solve the integrated vehicle and crew scheduling problem can potentially be improved” (Huisman, 2004, p. 149) and “we suggest to pursue further research on faster solution procedures for integrated problems” (Steinzen, 2007, p. 180)) and can be directly influenced by the transportation planning researchers themselves, whereas the former aspect cannot. It should therefore be noted that the evolution of technology (i.e. computers) is a real factor concerning the tractability of a specific transportation problem and thus forms an actual part of future research, although planning researchers have no direct impact on it. Because researchers (i.e. the authors of the considered papers) have no direct impact on it, the topic of improving computer performance was not listed. Moreover it should never be the standard procedure to just wait until there is a computer strong enough to solve a particular complex problem with an already existing algorithm38. That is not scientific progress, developing better algorithms that can solve the complex problem on an ordinary computer is. 38 Besides, some aspects of computation time are independent of the speed of the computer and thus also apply for even the most powerful computers. E.g. the computation time for solving a problem optimally can explode for a certain number of depots, if that number only slightly increases. (Huisman, 2004) 64 4. General conclusion In this thesis, we discussed the routing and scheduling of vehicles and crews, perhaps the most important problems faced in today’s transport companies. This thesis thus actually covered the whole of the planning process in transportation. Before starting with our actual exposition, we defined the year 2004 (more precisely January 1 st of this year) as the transition point between past and present, primarily because the central research question of this thesis, ‘Is it (always) advantageous to use an integrated approach for planning problems in transportation?’, was not considered until then. For the past, we provided the definitions for the individual routing and scheduling problems – the Vehicle Routing Problem (VRP), the Vehicle Scheduling Problem (VSP) and the Crew Scheduling Problem (CSP). In short, we could say that the VRP constructs routes so that a number of customers can be serviced with a fleet of vehicles, while the VSP assigns vehicles to cover these routes and the CSP allocates crews to operate the vehicles (and routes). The main objective is always to minimize total costs, taking into account certain constraints. We then situated these individual problems within the larger whole of the planning process in transportation companies, which is divided into a strategic, tactical, and operational phase. The routing (VRP) is mainly a strategic aspect, whereas the scheduling (VSP and CSP) is part of the operational phase. We also elaborated on the distinction between public transport (transporting passengers) and distribution context (delivering goods to customers), which was – just like the distinction between single and multiple depot problems – an important guideline throughout the thesis. In public transport, routes and timetables are mostly given (e.g. by a local authority) and remain unchanged for a long period of time, which is not the case for distribution companies. This led to the understanding that the combination of routing and scheduling is generally a concern only for those distribution companies, while in public transport the focus lies mainly on the mere scheduling of vehicles and crews. Planning in a distribution context thus comprises more aspects, but is certainly not intrinsically more difficult or more important than the public transport counterpart, where there is a much more delicate trade-off between customer and cost. The scheduling problem was described, first the traditional sequential approach (which basically seemed to come down to solving the CSP, because this always incorporates the solution of the VSP first) and then the integrated Vehicle and Crew Scheduling Problem or VCSP (including a brief pre-2004 literature review), defined in both public transport and distribution context to point out similarities and differences. An critical aspect is the possibility of introducing time windows in a distribution context, which is never done for public transport, where trips have to be punctual. 65 If the VCSP is about integration within the scheduling, the problem referred to as Vehicle and Crew Routing and Scheduling Problem (VCRSP) represents the strive for integration between routing and scheduling, and therefore only appears in a distribution context. It was proven that a VCRSP can be seen as some kind of merger between a VRP and a VCSP. The VCSP papers from the present were categorized according to a new procedure presented in this thesis. Of course, we first defined the categorization criteria, which either related to the practical problem considered or the solution method used for it. In descending order of importance, we identified 8 criteria of the first kind (type of transportation problem, mode of transportation, number of depots, objectives, size/practicality of the problem, degree of urbanization, regularity of the timetable, and admission of changeovers) and 5 of the second (degree of integration, degree of network segmentation, model, algorithm, and dynamism of the solution approach). The actual categorization was then performed for 18 of the most important VCSP approaches, along with the mentioning of a dozen other papers strongly related to them. We chose to only consider VCSP approaches for road traffic (the most discussed mode of transportation), supplemented by one paper on railways, but no airline problems. Extension of the categorization procedure for airlines and also for railways is thus an obvious working point. Some links between the proposed criteria were discovered and explained: · urban scenarios mostly correspond to single depot cases and ex-urban to multiple depots, · the service level objective (reducing delays) can be achieved by using a dynamic scheduling approach, and · the quality (regularity) of crew schedules is only a relevant objective when the timetable is irregular. Also several interesting trends were observed for the VCSP resulting from the categorization: · the vast majority of authors apply the proposed approaches to real-life problems, · the largest instance tackled so far is of the size of 1,414 trips and these sizes have not really increased in recent years whereas the number of depots and vehicle types considered did, · clearly more ex-urban scenarios are considered, · virtually all approaches make use of complete integration (as opposed to partial integration), · a set partitioning formulation is still the most popular one, and 66 · solution methods based on column generation remain the most commonly used with an evolution towards linear relaxation (as opposed to Lagrangian relaxation) to solve the master problem and branch-and-bound methods (as opposed to Lagrangian heuristics) to obtain feasible solutions. Most importantly, we showed that we can clearly and indisputably answer the VCSP-related part of the central research question, ‘Is it (always) advantageous to use an integrated approach for scheduling problems in transportation?’ with a sound ‘Yes’. For the description of recent developments of the VCRSP, we provided an overview of the existing approaches (7 relevant papers) and presented a first step towards an extensive categorization similar to that for the VCSP. A two-dimensional classification was introduced with dimensions corresponding to admission of changeovers and degree of integration, thus again identifying the two overarching categorization aspects of practical problem and solution method. The existing literature on the VCRSP is perhaps not yet extensive enough for an elaborate categorization to be of use at this time, but it is certainly an interesting possibility for the future. A couple of meaningful observations were made, namely that the concrete application contexts of the VCRSP are very varied (limousine rental, mail distribution, oil delivery, just to name a few) and that virtually every publication on the VCRSP so far uses a partially integrated approach (notice the contrast with the VCSP). The VCRSP-related part of the central research question, ‘Is it (always) advantageous to use an integrated approach for combined routing and scheduling problems in transportation?’, could not be as decisively answered as the VCSP-related part. We cannot just respond ‘Yes’, but nuance it by stating that there certainly is potential in the integration between routing and scheduling, at first sight less pronounced than for the VCSP, but in any case more research in the area of VCRSP is needed in order to evolve towards a universal conclusion. This further research can thereby lead to new methods which may well ever produce better results than the traditional approach, so we could then wholeheartedly answer ‘Yes’ to the central research question, just as for the VCSP. It was beyond the scope of the thesis to propose improvements for specific existing solution methods, more particular for models and algorithms. We rather identified more general and problem-inherent future research topics. These were primarily discovered by identifying gaps within the VCSP categorization and the VCRSP two-dimensional classification and then making recommendations for filling those gaps. When available, we also mentioned some quotes for each topic, literally extracted from the considered papers in order to substantiate the relevance of the topics. Some of the most important recommendations for the VCSP are: 67 · strive for the solution of the whole network at once for every transport company, · shift the attention a little from customer satisfaction (service level objective) towards wellbeing of employees (quality of crew schedules objective), · take into account the demand for more regular crew schedules when dealing with an irregular timetable, · keep pushing the size boundaries of instances, · consider multiple crew types instead of by default assuming only a single crew type, · investigate the possible advantage of allowing unrestricted changeovers (compared to restricted changeovers), · further research in the area of metaheuristics and constraint programming, and · pay more attention to the development of dynamic scheduling approaches (and more specifically introduce such an approach for the distribution context). It is remarkable that many authors seem to acknowledge the importance of the last mentioned topic. However, we observed that only very few papers actually present such a dynamic approach. We have thus identified a high need and a low presence, so consequently, dynamic scheduling should be treated as one of the hottest amongst the topics. We also came to the conclusion that it might be needed to dedicate some extra attention and effort to the whole of the VCRSP, since it is underdeveloped compared to the VCSP while nonetheless, distribution problems are an import aspect of today’s transport issues. Other significant VCRSP recommendations are: · develop and evaluate approaches pursuing complete integration, · test existing theoretical approaches on practical instances, · focus more on multiple depot problems where changeovers are not constricted to only take place at a single location, and · adapt VCRSP methods that are initially designed for one specific application in order to make them more generally applicable for other practical problems as well. The last topic is the most frequently mentioned research subject in other papers. This is of course not surprising, since it concerns a very general recommendation that actually reflects the logical lifecycle of a newly developed solution method. 68 Bibliography Ball, M., Bodin, L., Dial, R., 1983, A Matching Based Heuristic for Scheduling Mass Transit Crews and Vehicles, Transportation Science, 17, 4-31. Barnhart, C., Laporte, G., 2007, Transportation, Handbooks in Operations Research and Management Science, 14, Elsevier, Amsterdam. Bartodziej P., Derigs U., Malcherek D., Vogel, U., 2009, Models and algorithms for solving combined vehicle and crew scheduling problems with rest constraints: an application to road feeder service planning in air cargo transportation, OR Spectrum, 31, 405-429. Bertossi, A.A., Carraresi, P., Gallo, G., 1987, On Some Matching Problems Arising in Vehicle Scheduling Models, Networks, 17, 271-281. Bodin, L., Golden, B., Assad, A., Ball, M., 1983, Routing and scheduling of vehicles and crews: the state of the art, Computers & Operations Research, 10, 63-211. Borndörfer, R., Löbel, A., Weider, S., 2002, Integrierte Umlauf- und Dienstplanung im Nahverkehr, Technical Report ZR 02-10, Konrad-Zuse Zentrum, Berlin. Borndörfer, R., Löbel, A., Weider, S., 2004, A bundle method for integrated multi-depot vehicle and duty scheduling in public transit, Technical Report ZR 04-14, Konrad-Zuse Zentrum, Berlin. Bunte, S., Kliewer, N., 2006, An overview on vehicle scheduling models, Technical Report 11/2006, DS&OR Lab, University of Paderborn. Carpaneto, G., Dell'Amico, M., Fischetti, M., Toth, P., 1989, A branch and bound algorithm for the multiple depot vehicle scheduling problem, Networks, 19, 531-548. Carraresi, P., Gallo, G., 1984, Network models for vehicle and crew scheduling, European Journal of Operational Research, 16, 139-151. Cordeau, J., Stojkovíc, G., Soumis, F., Desrosiers, J., 2001, Benders Decomposition for Simultaneous Aircraft Routing and Crew Scheduling, Transportation Science, 35, 375-388. IX Darby-Dowman, K., Jachnik, J.K., Lewis, R.L., Mitra, G., 1988, Integrated decision support systems for urban transport scheduling: Discussion of implementation and experience. In Daduna, J.R., Wren, A., Proceedings of the Fourth International Workshop on Computer-Aided Transit Scheduling, Lecture Notes in Economics and Mathematical Systems, 226-239, Springer, Berlin. Desaulniers, G., Hickman, M.D., 2006, Public transit, In Barnhart, C., Laporte, G., Transportation, Handbooks in Operations Research and Management Science, 69-127, North Holland, The Netherlands. Drexl, M., Rieck, J., Sigl, T., Berning, B., 2011, Simultaneous Vehicle and Crew Routing and Scheduling for Partial and Full Load Long-Distance Road Transport, Technical Report LM-201105, Johannes Gutenberg University Mainz. Ernst, A.T., Jiang, H., Krishnamoorthy, M., 2004, Staff scheduling and rostering: A review of applications, methods and models, European Journal of Operational Research, 153, 3-27. Falkner, J., Ryan, D.M., 1992, Set partitioning for bus crew scheduling in Christchurch. In Desrochers, M., Rousseau, J., Computer-Aided Scheduling, Lecture Notes in Economics and Mathematical Systems, 386, 359-378, Springer, Berlin. Fischetti, M., Lodi, A., Martello, S., Toth, P., 1989, The fixed job schedule problem with workingtime constraints, Operations Research, 37, 395-403. Fischetti, M., Lodi, A., Martello, S., Toth, P., 2001, A polyhedral approach to simplified crew scheduling and vehicle scheduling problems, Management Science, 47, 833-850. Freling, R., Boender, C.G.E, Paixão, J.M. Pinto, 1995, An Integrated Approach to Vehicle and Crew Scheduling, Tech. rept. 9503/A, Econometric Institute, Erasmus University of Rotterdam. Freling, R., 1997, Models and Techniques for Integrating Vehicle and Crew Scheduling, PhD thesis, Erasmus University of Rotterdam. Freling, R., Wagelmans, A.P.M., Paixão, J.M. Pinto, 1999, An Overview of Models and Techniques for Integrating Vehicle and Crew Scheduling, In Wilson, N.H.M., Computer-Aided Transit Scheduling, 441-460, Springer, Berlin. X Freling, R., Paixão, J.M. Pinto, Wagelmans, A.P.M., 2001, Models and Algorithms for SingleDepot Vehicle Scheduling, Transportation Science, 35, 165-180. Freling, R., Huisman, D.,Wagelmans, A.P.M., 2003, Models and Algorithms for Integration of Vehicle and Crew Scheduling, Journal of Scheduling, 6, 63-85. Gaffi, A., Nonato, M., 1999, An Integrated Approach to Extra-Urban Crew and Vehicle Scheduling, In Wilson, N.H.M., Computer-Aided Transit Scheduling, 103-128, Springer, Berlin. Gintner, V., Steinzen, I., Suhl, L., 2006, A time-space network based approach for integrated vehicle and crew scheduling in public transport. In Binetti, M., Civitella, F., Liddo, E.D., Dell'Orco, M., Ottomanelli, M., Proceedings of the EWGT2006 Joint Conferences, 371-377. Gintner, V., 2007, Integrierte Umlauf- und Dienstplanung im ÖPNV, PhD thesis, University of Paderborn. Golden, B., Raghavan, S., Wasil, E., 2008, The Vehicle Routing Problem: Latest Advances and New Challenges, Operations Research/Computer Science Interfaces Series, 43, Springer, Berlin. Haase, K., Friberg, C., 1999, An Exact Branch and Cut Algorithm for the Vehicle and Crew Scheduling Problem. In Wilson, N.H.M., Computer-Aided Transit Scheduling, 63-80, Springer, Berlin. Haase, K., Desaulniers, G., Desrosiers, J., 2001, Simultaneous vehicle and crew scheduling in urban mass transit systems, Transportation Science, 35, 286-303. Hollis, B., Forbes, M., Douglas, B., 2006, Vehicle routing and crew scheduling for metropolitan mail distribution at Australia Post, European Journal of Operational Research, 173, 133-150. Hollis, B., 2011, Vehicle and Crew Routing and Scheduling, PhD thesis, University of Queensland. Huisman, D., Freling, R., Wagelmans, A.P.M., 2001, A Dynamic Approach to Vehicle Scheduling, Tech. rept. EI2001-17, Econometric Institute, Erasmus University of Rotterdam. XI Huisman, D., 2004, Integrated and Dynamic Vehicle and Crew Scheduling, PhD thesis, Erasmus University of Rotterdam. Huisman, D., Freling, R., Wagelmans, A.P.M., 2005, Multiple-Depot Integrated Vehicle and Crew Scheduling, Transportation Science, 39, 491-503. Huisman, D., Wagelmans, A.P.M., 2006, A solution approach for dynamic vehicle and crew scheduling, European Journal of Operational Research, 172, 453-471. Kéri, A., Haase, K., 2007, Simultaneous Vehicle and Crew Scheduling with Trip Shifting, Operations Research Proceedings, 467-472. Kim, B., Koo, J., Park, J., 2010, The combined manpower-vehicle routing problem for multistaged services, Expert Systems with Applications, 37, 8424-8431. Klabjan, D., Johnson, E., Nemhauser, G., Gelman, E., Ramaswamy, S., 2002, Airline crew scheduling with time windows and plane count constraints, Transportation Science, 36, 337348. Laurent, B., Hao, J., 2007, Simultaneous vehicle and driver scheduling: A case study in a limousine rental company, Computers & Industrial Engineering, 53, 542-558. Laurent, B., Hao, J., 2008, Simultaneous Vehicle and Crew Scheduling for Extra Urban Transports, Lecture Notes in Computer Science, 5027, 466-475. Löbel, A., 1997, Optimal Vehicle Scheduling in Public Transit, PhD thesis, Technische Universität Berlin. Mesquita, M., Paias, A., Respício, A., 2006, Branching approaches for the integrated vehicle and crew scheduling, Technical Report 9/2006, Operations Research Center (CIO), University of Lisbon. Mesquita, M., Paias, A., 2008, Set partitioning/covering-based approaches for the integrated vehicle and crew scheduling problem, Computers & Operations Research, 35, 1562-1575. XII Mesquita, M., Moz, M., Paias, A., Paixão, J., Pato, M., Respício, A., 2011, A new model for the integrated vehicle-crew-rostering problem and a computational study on rosters, Journal of Scheduling, 14, 319-334. Patrikalakis, I., Xerocostas, D., 1992, A new decomposition scheme of the urban public transport scheduling problem, In Desrochers, M., Rousseau, J.-M., Proceedings of the Fifth International Workshop on Computer-aided Scheduling of Public Transport, Lecture Notes in Economics and Mathematical Systems, 386, 407-425, Springer, Berlin. Pepin, A.-S., Desaulniers, G., Hertz, A., Huisman, D., 2006, Comparison of heuristic approaches for the multiple depot vehicle scheduling problem, Technical Report EI2006-34, Econometric Institute, Erasmus University of Rotterdam. Prescott-Gagnon, E., Desaulniers, G., Rousseau, L.-M., 2010, Heuristics for an Oil Delivery Vehicle Routing Problem, Les Cahiers du GERAD, 63. Rodrigues, M.M., de Souza, C., Moura, A., 2006, Vehicle and crew scheduling for urban bus lines, European Journal of Operational Research, 170, 844-862. Sandhu, R., Klabjan, D., 2007, Integrated Airline Fleeting and Crew-Pairing Decisions, Operations Research, 55, 439-456. Sato, T., Sakikawa, S., Morita, T., Ueki, N., Murata, T., 2009, Crew and Vehicle Rescheduling Based on a Network Flow Model and Its Application to a Railway Train Operation, Journal of Applied Mathematics, 39, 142-150. Schellinck, K., 2009, Ontwerp van een optimale geïntegreerde "Supplier Managed Inventory" en "Crew Scheduling" strategie: het routeringsaspect, master’s thesis, Ghent University. Scott, D., 1985, A Large Linear Programming Approach to the Public Transport Scheduling and Cost Model. In Rousseau, J.M., Computer Scheduling of Public Transport 2, 473-491, Elsevier, Amsterdam. Steinzen, I., 2007, Topics in integrated vehicle and crew scheduling in public transport, PhD thesis, University of Paderborn. XIII Steinzen, I., Suhl, L., Kliewer, N., 2009, Branching strategies to improve regularity of crew schedules in ex-urban public transit, OR Spectrum, 31, 727-743. Steinzen, I., Gintner, V., Suhl, L., Kliewer, N., 2010, A Time-Space Network Approach for the Integrated Vehicle- and Crew-Scheduling Problem with Multiple Depots, Transportation Science, 44, 367-382. Tosini, E., Vercellis, C., 1988, An interactive system for extra-urban vehicle and crew scheduling problems. In Daduna, J.R., Wren, A., Proceedings of the Fourth International Workshop on Computer-Aided Transit Scheduling, Lecture Notes in Economics and Mathematical Systems, 41-53, Springer, Berlin. Toth, P., Vigo, D., 2002, The Vehicle Routing Problem, SIAM Monographs on Discrete Mathematics and Applications, Philadelphia. Valouxis, C., Housos, E., 2002, Combined bus and driver scheduling, Computers & Operations Research, 29, 243-259. Xiang, Z., Chu, C., Chen, H., 2006, A fast heuristic for solving a large-scale static dial-a-ride problem under complex constraints, European Journal of Operational Research, 174, 11171139. Zäpfel, G., Bögl, M., 2008, Multi-period vehicle routing and crew scheduling with outsourcing options, International Journal of Production Economics, 113, 980-996. XIV Appendix A “We solve the traditional vehicle and crew scheduling problem, or in other words the CSP, in two steps. First we have to solve the vehicle scheduling problem (either the SDVSP or the MDVSP dependent of the number of depots), generate all feasible pieces and (optional) all feasible duties first. Then we select the optimal duties by solving the corresponding set covering model with a Lagrangian heuristic. In the remainder of this discussion we will assume that there is only one depot. However, the algorithm can be straightforwardly generalized to the multiple depot case. The solution of the SDVSP gives a set of vehicle blocks on which we can define the relief points. This results in a set of tasks from which we can easily enumerate all feasible pieces, since a piece is a feasible sequence of consecutive tasks on the same vehicle block only restricted by its duration. Next, we generate all feasible duties. Since duties often consist, in practice, of at most three pieces, we can do this by enumerating all possible combinations of pieces and check if such a combination is feasible. One may use a duty generation network, as proposed by (Freling, 1997), to generate duties consisting of more than three pieces. Finally, we select the duties with the Lagrangian heuristic, for which a global description is given here: Step 0: Initialization Generate a set of pieces such that each task can be covered by at least one piece. The initial set of columns consists of these pieces. Step 1: Computation of dual multipliers Solve a Lagrangian dual problem with the current set of columns. This yields a lower bound for the current set of columns. Step 2: Selection of additional columns Select columns from the previously generated duties with negative reduced cost. If no such columns exist, meaning that the lower bound computed in Step 1 is a lower bound for the overall problem (or another termination criterion is satisfied), go to Step 3; else, return to Step 1. Step 3: Construction of feasible solution Use all the columns selected in Step 0 and Step 2 to construct a feasible solution. We compute a feasible solution by solving a set covering problem in which we consider all the columns which have been selected along the way. We can either do this exactly, using a general or specialized integer programming solver, or heuristically.” (Huisman, 2004, pp. 32-33) For a more profound description of this method, we of course refer to (Huisman, 2004). A.1 Appendix B In general, we have a feasible vehicle schedule “if each trip is assigned to a vehicle, and each vehicle performs a feasible sequence of trips, where a sequence of trips is feasible if it is feasible for a vehicle to execute each pair of consecutive trips in the sequence. In the multiple depot case, we must also make sure that each trip is assigned to a vehicle from a depot that is allowed to drive this trip. The vehicle costs consist of a fixed component for every vehicle and variable costs for idle and travel time. It is allowed that a vehicle returns to a depot between two trips if there is enough time to do so. From a vehicle schedule it follows which trips have to be performed by the same vehicle and this defines so-called vehicle blocks. The blocks are subdivided at relief points, defined by location and time, where and when a change of driver may occur and drivers can enjoy their break. A task is defined by two consecutive relief points and represents the minimum portion of work that can be assigned to a crew. We distinguish between two types of tasks: trip tasks corresponding to (parts of) trips, and deadhead tasks of course corresponding to deadheading. A deadhead is a period that a vehicle is moving to or from the depot, or a period between two trips that a vehicle is outside of the depot (possibly moving without passengers). All trip tasks need to be covered by a crew, while the covering of deadhead tasks depends on the vehicle schedules and determines the compatibility between vehicle and crew schedules. In particular, each deadhead task needs to be assigned to a crew if and only if its corresponding deadhead is assigned to a vehicle. Note that more than one trip task may correspond to a single trip, depending on the relief points along that trip. Similarly, more than one deadhead task may correspond to a single deadhead. The tasks then have to be assigned to crew members. The tasks that are assigned to the same crew member define a crew duty. Together the duties constitute a crew schedule. We only consider the basic crew scheduling problem as discussed in section 1.1.3, where a schedule is feasible if each task is assigned to one duty, and each duty is a sequence of tasks that can be performed by a single crew, both from a physical and a legal point of view. In particular, each duty must satisfy several complicating constraints corresponding to work load regulations for crews. Typical examples of such constraints are maximum working time without a break, minimum break duration, maximum total working time, and maximum duration. These constraints can differ between different types of duties, e.g. early, split and late duties. The cost of a duty is usually a combination of fixed costs such as wages, and variable costs such as overtime payment. Finally, we define a piece (of work) as a sequence of tasks on one vehicle block without a break that can be performed by a single crew member without interruption. B.1 We also need to make some assumptions that will determine the modeling of the problem (i.e. the constraints in the model) and thus the complexity and veracity of the approach. An assumption that is almost always made (in the single depot case as well as in the multiple depot case), is that the cost function for the VCSP is the summation of the vehicle and crew scheduling cost functions, where the primary vehicle and crew scheduling objectives are to minimize the number of vehicles and crews, respectively. An additional term can be added to the cost function which corresponds to variable vehicle costs (e.g. distance travelled) and variable crew costs (e.g. overtime payment). Another common assumption is that the feasibility of a piece only depends on its duration, which is limited by a minimum and maximum piece length that follows from legal regulations. For the single depot case, it is also frequently assumed that all vehicles are available at all time and that they are identical. This because with these characteristics, the underlying vehicle scheduling problem is polynomially and thus rapidly solvable. The assumptions cited above hold in a lot of real-world applications arising from public transport scheduling.” (Huisman, 2004, pp. 38-39) Although, one can easily think of a situation where this will no longer be the case, for example in a transportation company that disposes of more than one vehicle type. For the MDVCSP, it is often assumed that “each vehicle and each crew member has its/his own depot, which respectively means that a vehicle starts and ends in the same depot (while the number of vehicles used per depot is unlimited) and that a duty of a single crew member has only tasks on vehicles from that depot (however, it is not necessary that every duty starts and ends in this depot).” (Huisman, 2004, p. 79) Of course, these assumptions impose restrictions on the associated models and algorithms. “Especially, the second assumption that all crew have their own depot and are only allowed to perform tasks on vehicles from their own depot, is a crucial one. If this assumption is not valid, the crew scheduling problem cannot be solved independently for each depot anymore. Moreover, in the integrated models and algorithms the constraints linking the vehicles and the crews for the different depots will not be independent anymore. Such a situation can occur in practice, where one can think of a situation where the driver first has a few tasks on a vehicle from his own depot and then on a vehicle from another depot, where he will be relieved at the end of his duty on a location close to his own depot. However, it is unlikely that such a situation occurs frequently, since not all trips are allowed to be assigned to each depot and if it is allowed, it is often possible to change vehicles between the depots.” (Huisman, 2004, p. 100) Other possible assumptions are that “there is continuous attendance, i.e. there is always a driver present if the vehicle is outside the depot (however, vehicle attendance at the depot is not necessary) and that changeovers, which is the change of vehicle of a driver during his break, are allowed. These last two assumptions imply that if a driver has no changeover, i.e. before B.2 and after the break he drives the same vehicle, there should be another driver on this vehicle during the break of the former driver, since there is otherwise nobody attending this vehicle.” (Huisman, 2004, pp. 79-80) Again, one can easily imagine a transport company where changeovers are not allowed, which means that the last assumption has to be altered and thus the associated model (i.e. the constraints) will be different. B.3 Appendix C “A task is characterized by a start and end location, a duration, a set of vehicle types that can be used to cover it (usually dictated by a minimum capacity requirement on the task, however there are often operational restrictions that further reduce the set of compatible vehicle types) and a time window, bounded by an earliest and latest starting time, within which it must be commenced. A vehicle schedule is a set of tasks connected by empty repositioning or deadhead trips as required, for the same vehicle type that starts and ends at the same depot and usually spans less than one 24 hour period. A crew schedule or duty describes the tasks undertaken by one driver for one shift. It is a collection of contiguous activities whose composition is governed by the tasks it covers, the vehicle schedules these tasks are assigned to and a set of regulations that usually vary for different types of drivers. A duty must start and end at the same depot. It begins with a sign on activity followed by a set of drive activities describing the tasks and deadheads covered, with an optional intervening break activity which must occur at one of a predefined set of break locations, usually the depots, not necessarily the home depot. A duty ends with a sign off activity. Deadhead trips may be needed in a duty for many reasons: to connect different tasks covered by the duty, to travel from(to) the depot to(from) the start(end) location of the first(last) task covered by the duty or to return to a depot to incorporate a break activity. Note that a driver must always drive a vehicle when deadheading. This is a characteristic that is unique to the distribution domain. In the public transport context drivers are usually allowed to move within the network on their own (by walking, catching a bus or plane, etc.). A special set of activities are associated with any vehicle changes (which usually take place at a depot) that occur in a duty. Before a vehicle may be driven as part of a duty, time is given to the driver to prepare the vehicle and before a driver finishes using a vehicle, time is given to return it. Thus, from a vehicle view point, a duty can be thought of as being composed of what are commonly called pieces of work. A piece of work is defined as a contiguous set of activities within the same duty, bounded by prepare and return activities, within which the same physical vehicle is being used. Note that we can choose whether or not drivers are only allowed to drive vehicles based at their home depot. A prescribed set of regulations exists for each duty type which govern the legality of a duty. This set of regulations includes a maximum duration, the maximum duration before and after a break activity and specific durations for sign on, prepare, break, return and sign off activities. A set of vehicle movements for all vehicle types between allowable locations are used to construct the deadheads that appear in duties. This set of vehicle movements obeys triangle inequalities for cost, distance and duration. The operational cost of a duty is a non decreasing function of the C.1 duration of the duty. It includes the drivers salary, which is proportional to the time worked, and the running costs associated with the vehicles used during the duty, which is the sum of the costs of the individual vehicle movements performed during the duty when covering tasks and deadheading. There is also a fixed cost associated with each vehicle and duty type.” (Hollis, 2011, pp. 41-42) C.2 Appendix D The following discussion is situated in a public transport context, although similar statements can be made for product delivery. Since we based our description of the planning process in section 1.2 on that for a bus company, we will now compare the known bus planning process with that for a railway and an airline operator, consecutively. So we start with the differences between railway planning and the planning process in a bus company. First of all, “timetabling in railway planning is much more complicated than in bus transport, since the infrastructure of a railway system has to be taken into account. For example, between each pair of trains on the same track there should be a minimum headway due to safety regulations. Furthermore, differences in speed should be taken into account: a fast Intercity service cannot start just after a slow train in the same direction. Similar remarks can be made about the railway stations, where for instance cross-platforms connections between certain trains are preferred. In a bus system all those complicating problems do not occur. The vehicle scheduling problem is also slightly different, since different train units can be combined to form one larger train. This is not possible with buses. The order of the units in a train is another complicating factor. Furthermore, other problems that do not exist in the bus context, like shunting at railway stations should be taken into account. Finally, the movement of empty trains is often restricted, which makes the problem in this aspect somewhat easier than bus scheduling, since there are much less possible solutions. The crew scheduling and rostering problem are very similar. There is only one major difference: in general, a train needs more crew members (driver and guards) than a bus (only a driver). However, this does not make the problem much more complicated, since drivers and guards can either be scheduled independently or they are scheduled as teams consisting of a driver and one or two guards. From these differences we can conclude that, in general, railway planning is more complicated than bus planning. Now for the differences between bus and airline planning, which are generally much smaller. For instance, since most airlines have a hub-and-spoke network, the timetabling step can be compared with the one in regional bus transport where connections should be ensured on the hubs. However, the number of slots (capacity) at the airports is limited for each airline. Vehicle scheduling is divided into two steps for most airlines: fleet assignment and aircraft routing. In the fleet assignment problem, types of planes are assigned to each flight, while in the D.1 aircraft routing problem the actual schedules are determined. The first problem has a planning horizon of a season and the second one of a week. This is the major difference with bus scheduling where the planning horizon is one day. On the other side, the trips in airline planning are much longer than in bus scheduling. Crew scheduling (and rostering) differ in the number of crew members that has to be scheduled. However, pilots, copilots, cabin personnel can be scheduled independently of each other, in general. A more important difference is that the crew related processes are mostly divided into three steps: duty scheduling, crew pairing and crew rostering. A pairing is defined here as a sequence of duties without a (long) rest period. There are also some differences from a computational point of view. Since the trips are much longer, the number of trips per duty is much smaller than for bus driver scheduling. Furthermore, crew is much more restricted to certain types of airplanes. Therefore, large problems can be split up easily. If we compare the differences above, it seems that airline and bus planning have the same degree of difficulty. This in contradiction with railway planning, which is much more complicated than bus and airline planning.” (Huisman, 2004, pp. 6-7) D.2 Appendix E E.1 (Huisman, 2004), Ch. 3.3 practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers public transport road traffic single depot cost reduction application-based RET, the mass transit company in Rotterdam maximum of 259 trips (1 depot, 1 vehicle type) urban RET operates within the city of Rotterdam regular timetable restricted changeovers a changeover is defined as the change of vehicle of a driver during his break solution method degree of integration degree of network segmentation complete integration each route separately at the RET, the whole planning process is solved line-byline, so the data corresponds to bus and driver scheduling of individual bus lines model set partitioning The mathematical formulation proposed for the VCSP is a combination of the quasi-assignment formulation for the vehicle scheduling problem, and the set partitioning formulation for crew scheduling, similar to (Freling et al., 2003). algorithm The author uses the relaxation of the proposed model, where all set partitioning type of constraints are first replaced by set covering constraints, which are subsequently relaxed in a Lagrangian way. Then the remaining Lagrangian subproblem can be solved by pricing out the duty variables (related to the crew) and solving a SDVSP for the trip variables (related to the vehicles). We can use the auction algorithm of (Freling et al., 2001) for solving the SDVSP. The author proposes a two phase procedure for the column generation pricing problem: in the first phase, a piece generation network is used to generate a set of pieces of work which serve as input for the second phase where duties are generated. At the end a feasible crew schedule is computed given the (feasible) vehicle schedule which resulted from solving the last Lagrangian subproblem. This is done by solving the CSP. dynamism of the solution approach static approach conclusion The results reported show that we can get good solutions within reasonable computation times on a personal computer. It was not expected that it would be useful to integrate under every circumstance. When changeovers are allowed in the test problems, one driver may be saved by considering an integrated approach. Of course, the saving of one driver is considerable when taking into account that the author investigated instances for one bus line each. The interpretation of the computational results depends on the ratio between the fixed vehicle and crew costs. If fixed vehicle costs are much higher as E.2 compared to crew costs it becomes less attractive to apply the integrated approach. On the other hand, if crew costs are higher the integrated approach becomes more attractive. The author also applied the integrated approach to vehicle and crew scheduling to the exact RET case. He handled complicating constraints, which have not been considered in the literature before. He also showed the results for two individual bus lines, including the RET bus line with the largest number of trips. For these lines the integrated problem could be solved in a reasonable amount of time, where the gap between the lower bound and the best feasible solution was less than 10% in all cases. The main conclusion is that we can save vehicles and/or crews by integrating the vehicle and crew scheduling problem, which may lead to a big decrease in costs. Another important result is that sometimes it is indeed possible to reduce the total costs by allowing changeovers more often (so by making the changeovers more unrestricted). E.3 (Huisman, 2004), Ch. 3.4 practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers solution method degree of integration degree of network segmentation public transport road traffic single depot cost reduction application-based RET, the mass transit company in Rotterdam maximum of 259 trips (1 depot, 1 vehicle type) urban RET operates within the city of Rotterdam regular timetable no changeovers complete integration each route separately at the RET, the whole planning process is solved line-byline, so the data corresponds to bus and driver scheduling of individual bus lines model set covering The VCSP without changeovers can be modeled in a straightforward way as a set covering problem due to the introduction of combined duties. A combined duty is defined as a feasible vehicle duty and one or more corresponding feasible crew duties. algorithm The algorithm here is almost similar to the one given in (Huisman, 2004), Ch. 3.3. Since the model here is a pure set covering model, Lagrangian relaxations can be obtained in a similar way as for a traditional CSP. The column generation pricing problem needs an additional procedure with respect to (Huisman, 2004), Ch. 3.3 in order to generate combined duties using previously generated crew duties as input. This results in a three phase procedure. The other difference with (Huisman, 2004), Ch. 3.3 is that we obtain a feasible solution in a similar way as for CSP by solving a set covering problem with the columns generated during the lower bound phase. So the difference is that we do not compute a feasible vehicle schedule first, since combined duties already define a vehicle schedule in itself. dynamism of the solution approach static approach conclusion The conclusion is the same as the one for (Huisman, 2004), Ch. 3.3, except for this addition. Based on the results we can conclude that the benefit of integration may be significant when changeovers are not allowed. In practice, it often occurs that either changeovers are not possible due to long distances or changeovers are not allowed for legal or technical reasons. Especially, in the case that (almost) no changeovers are allowed, integration is very attractive because more vehicles and/or crews can saved here. E.4 (Huisman, 2004), Ch. 4.2-4.3 related papers: (Huisman et al., 2005) practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers public transport road traffic multiple depot cost reduction application-based Conexxion, the largest bus company in the Netherlands maximum of 653 trips (4 depots39, 1 vehicle type) ex-urban regular timetable restricted changeovers definition of a changeover: the change of vehicle of a driver during his break solution method degree of integration degree of network segmentation complete integration part of the network The total set consists of 1,104 trips and 4 depots in the area between Rotterdam, Utrecht and Dordrecht, three large cities in the Netherlands. They consider 8 different problem instances for which the number of trips varies between 194 and 653 trips (< 1,104). model set partitioning It is an extension of the model in (Huisman, 2004), Ch. 3.3 to the multiple depot setting. algorithm The algorithm that is proposed to solve the presented model, is a combination of column generation and Lagrangian relaxation and is an extension of the algorithm proposed in (Huisman, 2004), Ch. 3.3.The main part of the algorithm is used to compute a lower bound and therefore the author uses a column generation algorithm. He solves the master problem with Lagrangian relaxation. Furthermore, he generates the duties in the column generation subproblem (pricing problem), using the two phase procedure of (Huisman, 2004), Ch. 3.3. Since there is no interaction between the different depots in the column generation subproblem, we can solve them separately for every depot. An extra phase comprising the deletion of columns is introduced, compared to the single depot case in (Huisman, 2004), Ch. 3.3. Finally, feasible solutions are computed using a Lagrangian heuristic. dynamism of the solution approach static approach conclusion The results reported indicate that medium-sized problem instances with multiple depots can be solved by using an integrated approach for the vehicle and crew scheduling problem. Furthermore, there are significant savings compared to the traditional sequential approach, where first the vehicle scheduling and afterwards the crew scheduling problem is solved. 39 However, not all trips were allowed to be driven by a vehicle from every depot, so the average number of depots a trip can be operated from was actually only 1.71 (<4). E.5 (Huisman, 2004), Ch. 4.4 related papers: (Huisman et al., 2005) practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers public transport road traffic multiple depot cost reduction application-based Conexxion, the largest bus company in the Netherlands maximum of 653 trips (4 depots40, 1 vehicle type) ex-urban regular timetable restricted changeovers definition of a changeover: the change of vehicle of a driver during his break solution method degree of integration degree of network segmentation complete integration part of the network The total set consists of 1,104 trips and 4 depots in the area between Rotterdam, Utrecht and Dordrecht, three large cities in the Netherlands. They consider 8 different problem instances for which the number of trips varies between 194 and 653 trips (< 1,104). model set partitioning The author proposes a mathematical formulation which has only variables related to crew duties. The vehicle schedule can be obtained implicitly from the crew schedule. This formulation can be derived from the one presented in (Huisman, 2004), Ch. 4.2-4.3, but is also equivalent to the formulation of (Haase et al., 2001) in the case of a single depot. The formulation is characterized by the adding of an extra decision variable to count the number of vehicles. algorithm The algorithm is again a Lagrangian heuristic based on column generation, similar to (Huisman, 2004), Ch. 4.2-4.3.The author proposes an algorithm that consists of two phases. In the first phase, a lower bound is computed using the proposed model by again combining column generation and Lagrangian relaxation. The author uses the columns generated during the first phase in the second one to find a feasible vehicle schedule and a corresponding crew schedule. The second phase is similar to the construction of feasible solutions described in (Huisman, 2004), Ch. 4.2-4.3. The important differences with (Huisman, 2004), Ch. 4.2-4.3 are thus in computing the lower bound, where the author uses a different model. dynamism of the solution approach static approach conclusion Same conclusions apply as for (Huisman, 2004), Ch. 4.2-4.3. However, the lower bounds obtained by this algorithm are rarely stronger than the bounds obtained by the one in (Huisman, 2004), Ch. 4.2-4.3 and 40 However, not all trips were allowed to be driven by a vehicle from every depot, so the average number of depots a trip can be operated from was actually only 1.71 (<4). E.6 regularly weaker. If the solutions of the (Huisman, 2004), Ch. 4.2-4.3 algorithm are compared with the ones of the algorithm here, it is difficult to conclude which one is better, since sometimes the first one gives the best solution and sometimes the second one. However, for larger random problem instances, the (Huisman, 2004), Ch. 4.2-4.3 algorithm performs better. E.7 (Huisman, 2004), Ch. 5.3 related papers: (Huisman & Wagelmans, 2006) practical problem type of transportation problem public transport mode of transportation road traffic number of depots multiple depot objectives cost reduction and service level size/practicality of the problem application-based Conexxion, the largest bus company in the Netherlands maximum of 304 trips (4 depots41, 1 vehicle type) degree of urbanization ex-urban regularity of the timetable regular timetable admission of changeovers restricted changeovers changeovers only occur during the break solution method degree of integration complete integration degree of network segmentation part of the network The total set consists of 1,104 trips and 4 depots in the area between Rotterdam, Utrecht and Dordrecht, three large cities in the Netherlands. The considered problems consist of 164 and 304 trips (< 1,104). model set partitioning The mathematical formulation is based on that of (Huisman, 2004), Ch. 3.3. The notation with respect to the vehicle scheduling part of the formulation is completely similar to (Huisman et al., 2001). An important feature is that different scenarios for the travel times can be introduced. Hereby, the author assumes that one of the scenarios is the main scenario and that the variables related to the crew scheduling part of the problem are defined on this scenario. algorithm In the case of multiple-depots, all algorithms use the cluster-reschedule heuristic, so that multiple depot problems can be solved as several single-depot problems (the actual algorithm thus solely focuses on the single depot case). The algorithm solves a sequence of integrated vehicle and crew scheduling problems and is based on that in (Huisman, 2004), Ch. 3.3. The generation of columns is exactly the same since another vehicle scheduling problem does not influence this. The main difference lies in the approach for the crew-related constraints which are first replaced by set covering constraints and subsequently relaxed in a Lagrangian way. At the end a feasible crew schedule is computed given the (feasible) vehicle schedule for the main scenario which resulted from solving the last Lagrangian subproblem. dynamism of the solution approach dynamic approach conclusion The extension of the dynamic vehicle scheduling problem of (Huisman et al., 2001) to the situation where crews are also considered did not always give the results which was expected beforehand. For the small instance with a single depot the dynamic approach performs well. However, computation 41 However, not all trips were allowed to be driven by a vehicle from every depot, so the average number of depots a trip can be operated from was actually only 1.71 (<4). E.8 times are still high for applying such an approach in practice. On the other hand for the medium-sized instance with multiple depots, the traditional static approach with buffer times performed much better. The first reason is that this approach used the cluster-reschedule heuristic, i.e. all trips are assigned beforehand to a certain depot. In other words, the overall result is dependent on the chosen assignment. Another reason could be that the computation times allowed to solve the approach dynamically were set too small, although by extending these times the results did not improve so much. The final mentioned reason is that the idea of dynamically solving itself does not work so well. Since the dynamic approach worked well for the small instance where the data set was not divided into several smaller ones, and since extending the computation times did not lead to significant improvement, we can conclude that the way the problem is split up, is the bottleneck. Therefore, the author recommends to invest further research in speeding up the suggested algorithms. With faster computers and better algorithms the dynamic approach should outperform the static one with buffer times for larger problem instances as well. Finally, the author makes several remarks about the practical applicability of such a dynamic approach. First of all, it will be difficult to test the assumptions that were made in a practical environment. For instance, how can one measure if the travel times of the trips are really independent of the actual chosen schedule? Secondly, it is important how drivers (but also planners and managers) react on such a way of working. It is very easy for them to frustrate such an approach. Therefore, the author concludes the discussion with the fact that there is still a long way to go before such an approach can be used in practice. E.9 (Borndörfer et al., 2004) practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers solution method degree of integration degree of network segmentation public transport road traffic multiple depot cost reduction application-based the Regensburger Verkehrsbetriebe GmbH (RVB), a medium sized public transportation company in Germany and the Regionalverkehrsbetrieb Kurhessen (RKH), a regional carrier in the middle of Germany maximum of 1,414 trips (1 depot42, 3 vehicle types) urban/ex-urban the integrated scheduling method is applied to halfregional, half-urban instances and other mainly regional instances irregular timetable the RVB operation has a different scenario for Sundays and for workdays restricted changeovers complete integration whole network at once the authors consider instances that contain the entire RVB operation model multicommodity flow and set partitioning The model consists of a multicommodity flow model for vehicle scheduling and a set partitioning model for duty scheduling on timetabled trips. These two models are joined by a set of coupling constraints for the deadhead trips. algorithm The authors use a Lagrangian relaxation approach to solve the proposed model. Relaxing the coupling constraints results in a Lagrangian master problem, a vehicle scheduling and a duty scheduling problem. They use the method of (Löbel, 1997) to solve the vehicle scheduling problem, column generation to solve the duty scheduling problem and an inexact adaptation of the proximal bundle method to solve the Lagrangian master problem, producing dual and additional primal information as opposed to a subgradient algorithm. After computing a lower bound, the bundle core is called repeatedly in a branchand-bound type procedure (backtracking procedure) to produce integer solutions. dynamism of the solution approach static approach conclusion The authors show that it is possible to tackle large-scale, complex, real-world integrated vehicle and duty scheduling problems. The largest and most complex instance up to now has that been attacked with integrated scheduling techniques is solved in about 125 hours. Furthermore, they compare their 42 For another instance, 3 depots were considered with 634 trips and 5 vehicle types. E.10 approach with (Huisman, 2004), Ch. 4.2-4.3 on the same set of artificial instances. Using the same assumptions, their approach clearly outperforms Huisman's method and solves instances with up to 400 trips and 2 (4) depots in 3.3 (12) hours. Also, the solutions produced can be decidedly better in several respects at once than the results of various types of sequential planning. E.11 (Rodrigues et al., 2006) practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers public transport road traffic single depot cost reduction application-based three companies that operate in the large metropolitan regions of São Paulo and São Bernardo do Campo, Brazil maximum of 395 trips (1 depot, 1 vehicle type) urban regular timetable passenger demand on a typical day is considered no changeovers a crew is designated to a single vehicle for the entire duration of its daily work schedule solution method degree of integration degree of network segmentation complete integration each route separately 7 different lines are considered model set covering/set packing The techniques used are based on integer programming models. algorithm A hybrid strategy combining mathematical programming models and heuristics is proposed. The former produce good feasible solutions, while the latter improve the quality of the final solutions. The algorithm has four phases: a preliminary schedule generator, a vehicle block generator, a final schedule generator, and a heuristics that adjusts trip departure times. In the first phase, a bipartite graph is constructed in order to obtain a set of primary start times. In the next phase, the primary start times are used to generate vehicle blocks. In the third phase, the vehicle blocks are used in a classical packing or covering model in order to construct the schedule. After once or twice through the cycle involving the first three phases, the fourth and final phase starts. Here, a simple greedy heuristic adjusts the trip departure times that are still not adequately well spaced by the end of the cycle. dynamism of the solution approach static approach conclusion This hybrid strategy was able to produce quite adequate solutions, in a fraction of the time that experts take to construct manual solutions. Also, the operational cost of the proposed method showed considerable gains over the manual scheduling approach. E.12 (Steinzen, 2007), Ch. 2.3-2.4 related papers: (Gintner, 2007) practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers solution method degree of integration degree of network segmentation public transport road traffic multiple depot cost reduction application-based Conexxion, the largest bus company in the Netherlands maximum of 653 trips (4 depots43, 1 vehicle type) ex-urban regular timetable restricted changeovers complete integration part of the network The total set consists of 1,104 trips and 4 depots in the area between Rotterdam, Utrecht and Dordrecht, three large cities in the Netherlands. They consider 8 different problem instances for which the number of trips varies between 194 and 653 trips (< 1,104). model multicommodity flow and set partitioning The formulation is introduced by (Gintner, 2007) and combines a multicommodity network flow formulation for vehicle scheduling with a set partitioning formulation for crew scheduling. The underlying vehicle scheduling network is structured as a time-space network44. algorithm The solution method is a combination of column generation and Lagrangian relaxation and has been inspired by (Huisman, 2004), Ch. 4.2-4.3. More precisely, column generation is used to compute a lower bound where Lagrangian relaxation is applied to solve the master problem. Instead of solving the restricted master problem with the simplex method to optimality, a subgradient method is used to solve the Lagrangian dual approximately. Then, the two phase pricing procedure for the column generation pricing problem proposed in (Huisman, 2004), Ch. 3.3 is used. The final step of the solution method aims at finding a pair of feasible and compatible vehicle and crew schedules with a Lagrangian heuristic. dynamism of the solution approach static approach conclusion This method was not tested in itself, because it was improved within the same paper, namely in (Steinzen, 2007), Ch. 3.1-3.3. Therefore, we refer to the conclusion for that approach. 43 Here a different setting (compared to (Huisman, 2004), Ch. 4.2-4.3) is considered, where every trip may be serviced from every depot. Obviously, this makes the problems more difficult since the solution space is expanded. 44 Multicommodity network flow formulations for multiple depot vehicle scheduling problems can be classified by the underlying network structure. In a connection-based network (CBN), each feasible connection between two trips corresponds to an explicit arc in the network while in a time-space network (TSN) only connections between groups of compatible trips are considered. A time-space network approach reduces the number of connection arcs dramatically if the number of start and end locations is small compared to the number of trips. (Steinzen, 2007) E.13 (Steinzen, 2007), Ch. 3.1-3.3 related papers: (Steinzen et al., 2010) practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers solution method degree of integration degree of network segmentation public transport road traffic multiple depot cost reduction application-based Conexxion, the largest bus company in the Netherlands maximum of 653 trips (4 depots45, 1 vehicle type) ex-urban regular timetable restricted changeovers complete integration part of the network The total set consists of 1,104 trips and 4 depots in the area between Rotterdam, Utrecht and Dordrecht, three large cities in the Netherlands. They consider 8 different problem instances for which the number of trips varies between 194 and 653 trips (< 1,104). model multicommodity flow and set partitioning The model is the same as in (Steinzen, 2007), Ch. 2.3-2.4. algorithm The solution approach is based on Lagrangian relaxation in combination with column generation. Again, column generation is used to compute a lower bound where Lagrangian relaxation is applied to solve the master problem. The two phase pricing procedure for the column generation pricing problem of (Huisman, 2004), Ch. 3.3 is used, but now with a novel time-space network46 (instead of a classic connetion-based network) for the duty generation phase. Also, a new dynamic programming approach (including labeling approaches) to the resource constrained shortest path problems that appear in the duty generation phase is proposed. The performance of the standard version of the method is then considerably improved by using preprocessing (both generic and problem-specific) and further acceleration techniques (multiple pricing, restricted networks, state space reduction and label pruning). The author discusses three methods to compute integer solutions: a Lagrangian heuristic (sequential approach), a branch-and-bound approach with novel branching schemes (branching on variables and branching on follow-ons), and a novel heuristic branch-and-price algorithm (fix-and-optimize) which enhances the method of (Huisman, 2004), Ch. 4.2-4.3 by regenerating columns in the integer phase and applying depth-first (heuristic) branching in combination with different fixing strategies (fixing service trips/follow-ons to depots). The latter approach proved to be the best and was therefore used. dynamism of the solution approach static approach 45 Here a different setting (compared to (Huisman, 2004), Ch. 4.2-4.3) is considered, where every trip may be serviced from every depot. Obviously, this makes the problems more difficult since the solution space is expanded. 46 The author also presented a similar aggregated time-space network, but since this network resulted in rather poor solutions it was immediately discarded in the paper, as in this thesis. E.14 conclusion Notice that this setting is different to the results published in (Huisman, 2004), Ch. 4.2-4.3 and (Huisman et al., 2005) since the solution space is expanded. As a consequence, the results cannot be directly compared. The results show that real-world instances with up to 653 trips and 4 depots can be solved. Furthermore, there is an efficiency gain compared to sequential planning. It can be observed that the computational time does not always increase with the problem size. The author concludes that his algorithm performs better if the density of the columns is small. The author reports a strong impact of column density on the computational burden of a column generation algorithm for a multiple depot vehicle scheduling problem. Finally, the number of vehicles is always minimal for the presented approach, i.e., equals the number of vehicles when sequential planning is performed. The approach was also applied to randomly generated data instances. Similar to the results on real-world problem instances, the total number of vehicles and drivers can be remarkably reduced if integrated planning is performed. It is shown that our approach clearly outperforms all other approaches from literature ((Gintner et al., 2006), (Borndörfer et al., 2004) and (Huisman et al., 2005)) in terms of solution quality and solution time. Furthermore, the author has so far tackled the largest instances with 4 or more depots. Also, approaches based on models like (Steinzen, 2007), Ch. 2.3-2.4 (also used by (Gintner et al., 2006)) are beneficial compared to the classic connection-based model of (Huisman, 2004), Ch. 4.2-4.3. The presented approach requires between 29% and 73% of the computational time of (Borndörfer et al., 2004). Furthermore, the results indicate that the approach is the overall fastest known method for integrated vehicle and crew scheduling problems under the stated assumptions. The results indicate that the proposed approach can efficiently cover duty types with many pieces of work and complex feasibility rules. E.15 (Steinzen, 2007), Ch. 3.4 practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers solution method degree of integration degree of network segmentation public transport road traffic multiple depot cost reduction application-based Conexxion, the largest bus company in the Netherlands maximum of 653 trips (4 depots47, 1 vehicle type) ex-urban regular timetable unrestricted changeovers complete integration part of the network The total set consists of 1,104 trips and 4 depots in the area between Rotterdam, Utrecht and Dordrecht, three large cities in the Netherlands. We consider 8 different problem instances for which the number of trips varies between 194 and 653 trips (< 1,104). model multicommodity flow and set partitioning The model is very similar to that of (Steinzen, 2007), Ch. 2.3-2.4. The main difference is that, if we allow unrestricted changeovers, drivers may use vehicles from all depots. So that the duty variables and linking constraints, which are separated by depot in (Steinzen, 2007), Ch. 2.3-2.4, do not need to be separated anymore here. algorithm Similar to the solution approach in (Steinzen, 2007), Ch. 2.3-2.4 column generation in combination with Lagrangian relaxation is applied. Basically, the same constraints are relaxed in a Lagrangian way and again the same two phase pricing procedure is used as proposed in (Huisman, 2004), Ch. 3.3. However, we no longer have a separate pricing problem for each depot since in the proposed model the duty variables are not separated by depot. Consequently, the author sets up a single piece generation network, which is an acyclic directed time-space network. Feasible solutions are found by using a heuristic branch-and-price approach similar to (Steinzen, 2007), Ch. 3.1-3.3. dynamism of the solution approach static approach conclusion Basically, the results show that there is an efficiency gain if vehicle and crew scheduling are treated in an integrated way. Similar to the restricted case, most of the time is spent in the integer phase. The author concludes that the model with unrestricted changeovers is computationally more attractive than that with restricted changeovers. Furthermore, he believes it is worthwhile for planners in practice to allow unrestricted changeovers since the additional exibility results in efficiency gains. The results show 47 Here a different setting (compared to (Huisman, 2004), Ch. 4.2-4.3) is considered, where every trip may be serviced from every depot. Obviously, this makes the problems more difficult since the solution space is expanded. E.16 that the presented approach outperforms the method of (Mesquita & Paias, 2008) in terms of computational time and/or solution quality. Furthermore, instances with 640 trips and 4 depots are solved that have not been tackled before. Finally, the author mentions that a valid lower bound can be computed with his method while this is not possible for the method of (Mesquita & Paias, 2008) (since they heuristically define the set of tasks). E.17 (Steinzen, 2007), Ch. 4 practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers solution method degree of integration public transport road traffic multiple depot cost reduction theoretical randomly generated instances48 maximum of 200 trips (4 depots, 1 vehicle type) ex-urban regular timetable restricted changeovers complete integration although fully integrated fitness evaluation is only performed for the first few iterations (after that, a traditional sequential approach is used), it can be used throughout the entire evaluation process as well degree of network segmentation n/a model multicommodity flow and set partitioning The model is initially the same as (Steinzen, 2007), Ch. 2.3-2.4, but it is decomposed into different subproblems. algorithm The author presents a novel hybrid evolutionary algorithm that combines mathematical programming techniques with an evolutionary algorithm49. The algorithm is based on a problem decomposition that first assigns trips to depots (providing a trip-depot vector) and thus reduces the multiple-depot integrated problem to several integrated problems with a single depot. The EA is used to find a good trip-depot assignment where the fitness of a chromosome (individual) is evaluated using mathematical programming techniques. In particular, the author uses column generation in combination with Lagrangian relaxation. The computation of the fitness of the individuals can be done in three different ways: sequential, partially integrated, or fully integrated. When we use a fully integrated evaluation, then the model of (Steinzen, 2007), Ch. 2.3-2.4, for a given trip-depot assignment, reduces to a minimum cost flow problem in combination with a set partitioning problem. Then column generation is used in combination with Lagrangian relaxation in a similar way as in (Steinzen, 2007), Ch. 2.3-2.4. dynamism of the solution approach static approach conclusion The results reported in the previous section indicate that medium-sized problem instances with multiple 48 Proposed by (Huisman, 2004). “An Evolutionary Algorithm (EA) simulates evolutionary processes in nature by creating an initial population of individuals and applying genetic operators in each generation/reproduction. Each individual is represented by a string or chromosome and corresponds to a possible solution to the (combinatorial) optimization problem. The fitness of an individual represents the value of the objective function. Furthermore, individuals with a high fitness get the opportunity to reproduce among each other by exchanging genetic information.” (Steinzen, 2007, p. 125) 49 E.18 depots can be solved by using the proposed evolutionary algorithm. Furthermore, the approach discloses significant savings compared to the traditional sequential approach without requiring a fully integrated solution method. Although the presented algorithm performs worse than the best known integrated algorithm, it proved to be competitive with other integrated approaches from literature especially for medium-sized instances. E.19 (Steinzen, 2007), Ch. 6 related papers: (Steinzen et al., 2009) practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers solution method degree of integration public transport road traffic single depot cost reduction and quality of crew schedules application-based not mentioned which particular real-world problem maximum of 433 trips (1 depot, 1 vehicle type) ex-urban irregular timetable restricted changeovers partial integration (crew first - vehicle second) the Independent Crew Scheduling Problem (ICSP) is solved first and, then, the vehicle rotations from the crew scheduling solution are put together such that the vehicle schedule is feasible degree of network segmentation n/a not mentioned which particular real-world problem is considered, so not applicable model set covering The ICSP can be formulated as set covering problem. algorithm First, the ICSP is solved using a column generation algorithm in combination with Lagrangian relaxation. The author solves the corresponding Lagrangian dual with a subgradient algorithm to obtain approximate dual values. The column generation pricing problem corresponds to a resource constrained shortest path problem and is solved with a dynamic programming algorithm. The columns generated in the column generation phase serve as input to the second phase where an appropriate integer solution is sought. The author suggests a method for the second phase that takes the trade-off between costs and regularity into account (so the regularity objective is taken into account in the feasible solution construction phase). The basic idea of this method is to systematically search an optimal solution among all optimal solutions that is as similar as possible to a given reference solution. In particular, local branching cuts are used to select suitable solution subspaces and explore these subspaces with an adapted version of follow-on branching. Once the ICSP is solved, the vehicle rotations from the crew scheduling solution are put together such that the vehicle schedule is feasible. dynamism of the solution approach static approach conclusion A computational study that involved randomly generated and real-life data showed the applicability of the proposed techniques. The author concludes that local branching effectively improves the regularity while a carefully chosen follow-on branching scheme is well suited to improve solution quality and time. The combination of both methods leads to improved solutions in terms of both cost and regularity compared to a traditional approach with a default branch-and-bound approach. E.20 (Kéri & Haase, 2007) practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable public transport road traffic single depot cost reduction and service level loss of service quality can be avoided: the waiting times of passengers does not change if the trips of the lines whose connection are important are in the same flexible groups theoretical two smaller artificial test instances maximum of 133 trips (1 depot, 1 vehicle type) urban irregular timetable a flexible timetable where starting times of the trips can be shifted, will not be regular (not the same for each day) no changeovers admission of changeovers solution method degree of integration complete integration degree of network segmentation n/a model set partitioning The authors extend the model introduced in (Haase et al., 2001) to incorporate what is called trip shifting, where each trip is assigned an allowable set of shifted starting times. algorithm A heuristic method is proposed, nevertheless based on column generation. The linear relaxation of the proposed model represents the main problem of the column generation process. In each iteration of the process this relaxed problem is solved to optimality with the actual columns. Solving the subproblems is equivalent to finding a resource constrained shortest path in each driver network. The authors use an adapted dynamic programming algorithm (labeling algorithm) to solve this. After finding the optimal solution of the linear relaxation of the proposed model, a round-up method is used to achieve an integer solution. dynamism of the solution approach static approach although not explicitly performed in the paper, it should be logical (and possible – regarding the computation time – for very small instances) to use the proposed approach in a dynamic environment conclusion The authors have run the test first without using flexible timetable (but still using the same algorithm), then with using it. The new approach yields a much better solution regarding the number of buses, and the crew cost. However it requires much more time to solve the problem. E.21 (Mesquita & Paias, 2008) related papers: (Mesquita et al., 2006) and (Mesquita et al., 2011)50 practical problem type of transportation problem public transport mode of transportation road traffic number of depots multiple depot objectives cost reduction size/practicality of the problem theoretical randomly generated data problems51 maximum of 400 trips (4 depots, 1 vehicle type) degree of urbanization ex-urban not explicitly mentioned, but inferred from the test instances used regularity of the timetable regular timetable admission of changeovers unrestricted changeovers A break can occur during a changeover (but does not have to). After taking his break (or not), the driver may pick the same or another vehicle from any depot. solution method degree of integration complete integration degree of network segmentation n/a model multicommodity flow and mixed set partitioning/covering The VCSP is described by an integer linear programming formulation combining a multicommodity network flow model for the vehicle scheduling with a mixed set partitioning/covering model for the crew scheduling. algorithm The authors propose an algorithm that starts with a pre-processing phase, based on the optimal solution of the vehicle scheduling problem without requiring that vehicles return to the source depot, to define the set of tasks and to obtain an initial set of duties. In a second phase, they solve the linear programming relaxation of the models using a column generation scheme. The columns corresponding to the duties can be seen as paths in an adequate network and are generated as needed by solving shortest path problems with resource constraints (pricing problem). The pricing problem is solved by a heuristic procedure using dynamic programming and a reduced state space where states are associated to crew duties and the stages to tasks. Whenever the resulting solution is not integer, branch-andbound techniques are used over a subset of feasible duties for the crews. dynamism of the solution approach static approach conclusion Regarding results in (Huisman et al., 2005) and (Borndörfer et al., 2004) for the same test instances, the presented approach led to a smaller number of crews although, in some cases, a greater number of vehicles. However, better values for the sum of vehicles and crews were obtained, so a better general 50 Although (Mesquita et al., 2011) also comprises crew rostering within the operational planning phase, in addition to vehicle and crew scheduling. 51 Proposed by (Huisman, 2004). E.22 quality of the solution. The authors think that an important improvement over the existing methods is the time consumed by the proposed algorithm to obtain these results. They cannot make a direct comparison, since different computers have been used by the different authors. However, they state that when the size of the problem increases, the time spent by the presented algorithm becomes significantly smaller than the time spent by the algorithm proposed in (Borndörfer et al., 2004). From a transportation company point of view, it is an important feature of an algorithm to produce quick and ‘good’ solutions. Moreover, the resulting solutions have few over-covers and are similar to partitiontype solutions and this makes them easier to implement in a real situation. In conclusion, the proposed method seems to be a promising tool for dealing with large instances of the integrated VCSP. E.23 (Laurent & Hao, 2008) practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers solution method degree of integration degree of network segmentation public transport road traffic single depot cost reduction application-based 7 non-specified real-world instances maximum of 249 trips (1 depot, 2 vehicle types) ex-urban regular timetable no changeovers complete integration n/a it is not mentioned which real-life network was considered, although the maximum size of the instances gives away that the network was not tackled as a whole model constraint-based The authors introduce an original formulation relying on a constraint satisfaction and optimization model. This constraint-based formulation offers a natural modeling of the initial problem and provides a flexible basis to implement various metaheuristics. algorithm They present the first application of a metaheuristic, Greedy Randomized Adaptive Search Procedure (GRASP), to the VCSP. Within the GRASP algorithm, constraint programming techniques are first used to build initial solutions. Improvements of these solutions are achieved with a local search algorithm which embeds a powerful ‘ejection chain’ neighborhood exploration mechanism. dynamism of the solution approach static approach conclusion First, one observes that the integrated approach always outperforms the sequential one, or at least furnishes equivalent results. In particular, the savings in terms of number of drivers are significant. The sequential approach provides a lower bound for the number of vehicles that is always reached in the integrated solutions. Across the 7 instances, the results are also quite stable with very small standard deviations. Second, the integrated approach is more powerful than the sequential one in the sense that the sequential approach failed to solve a particular instance where the integrated approach succeeded. Moreover, in some cases, the integrated approach is indispensable, especially when relief opportunities are rare. These results show the dominance of the integrated approach over the sequential one. E.24 (Bartodziej et al., 2009) practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers distribution road traffic more specific a Road Feeder Service52 single depot trips starting and ending at the central hub cost reduction application-based a major German RFS-carrier maximum of about 1,400 trips (1 depot, number of vehicle types not mentioned53) ex-urban very large distances need to be traveled irregular timetable during operation of the fixed timetable, airlines will eventually ask the trucking company for additional transportation tasks on the spot, so-called ad-hocs no changeovers Drivers stick with their truck. Trips with two drivers are possible, but a relief in this context does not correspond to a changeover as defined, because the relieved driver does not switch to another vehicle but stays in the same truck. In fact, the two drivers could be seen as an entity. solution method degree of integration degree of network segmentation complete integration whole network at once The largest instance, representing a set of lines in a bidding round, has to be evaluated on the strategic planning level. Since design of the network is included here, of course a whole network is considered. model set partitioning The problem can be represented by a set partitioning type of formulation. algorithm two different algorithms were presented: 1) The authors describe an algorithm for obtaining near-optimal solutions for the proposed model by solving an LP relaxation via column generation and by using the generated columns to construct a feasible integer solution. In their implementation they use two complex compatibility graphs to represent all feasible round trips (columns). These graphs are constructed in a first pre-processing phase. 52 A Road Feeder Service (RFS) concerns the part of the air cargo transport process that is done over ground by a trucking company. 53 Another instance considers no less than 27 vehicle types for 779 trips. E.25 2) The authors also describe several local search-based metaheuristics for solving the proposed model. They combine complex and problem specific operations for the successive improvement of an initial solution from two different classes: block operations (which remove one or several blocks from the round trips of the current solution, combine the associated trips to new blocks and finally reassign each new block to a round trip/vehicle observing all constraints) and trip operations (which remove and reinsert single trips from/to the current solution, respectively). In each iteration of the local search improvement phase a type of neighborhood (two where defined) is selected randomly and applied to the current solution. Then, depending on the specific metaheuristic criterion (Simulated Annealing (SA), Great Deluge Algorithm (GDA) or Record-to-Record Travel (RRT)) the modified solution is accepted or rejected. dynamism of the solution approach static approach conclusion On the small examples (< 167 trips), the solution of the LP relaxation was optimal in most of the cases and the gap to the optimal integer solution was very small for the remaining instances. Yet, it was not possible to apply CG to the set of large instances (> 695 trips). As can be seen from the results the limiting factor is not computational time but too high in-core memory requirements. In order to solve the large instances, one must resort to the proposed metaheuristics. SA, GDA and RRT show a relatively similar convergence. The mean percentage deviation from the near-optimal integer CG-solutions (for the small instances) is for all metaheuristics smaller than 3% after 3 min running time. The authors state that RRT has the fastest convergence at the beginning and is only caught up if the running times are relatively large. This is probably due to the fact that the RRT-control is directly connected to the quality of the solution. On the other hand they report that when the first objective is to minimize the number of auxiliary vehicles SA and GDA are slightly preferable. Finally, RRT was applied to the largest instances. It took about 3 min to construct an initial feasible solution. Then, within 10 s only the initial solution could be improved by 24% and at the end of running time an improvement of 33% was gained. After all, the model and the heuristics have shown to be appropriate to be implemented in a decision support system for RFS-planning. E.26 (Sato et al., 2009) practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers public transport railway multiple depot cost reduction and service level prevent delays from escalating is an objective (i.e. equal to the service level objective) application-based a not explicitly named Japanese railway line maximum of 786 trips (exact number of depots or vehicle types was not mentioned) ex-urban railway trains operating between multiple cities are considered (no metros or trams) regular timetable the line has about 200 vehicles and approximately 800 trains that are operated on the line everyday restricted changeovers there are certain resources (drivers, conductors) which cannot be allocated to certain trips for some reason, such as vehicle type solution method degree of integration degree of network segmentation complete integration each route separately the Japanese railway lines form a huge and complicated traffic network because of their interconnectedness, a single railway line is considered model multicommodity flow A 0-1 integer programming formulation based on a network flow is proposed. algorithm The authors propose the following two-phase solution approach. In the first phase a feasible solution is generated by using a partial exchange (is the exchange of a part of a schedule with another one, performed when some flows have become infeasible because of transport disruption and timetable changes) as a heuristic flow modification to make a feasible schedule. After setting this solution as an initial solution, the solution method searches for alternatives by a local search in the second phase. Local search is a kind of generate-and-test method in which a neighborhood of the temporal solution is generated at each iteration step. This gives a set of solutions similar to the temporal solution, and the solution with the best evaluation value is selected as the improved temporal solution. dynamism of the solution approach dynamic approach a method for rescheduling is of course dynamic conclusion The proposed formulation is able to represent the differences between the new and original schedules, E.27 which is a significant criterion for the rescheduling problem, though taking these differences into account is difficult for other related formulations based on the set partitioning/covering models. Computational results of real-world vehicle rescheduling data from the railway line indicated that the proposed method generated a feasible solution within a practical amount of time, and on the basis of a two-phase solution approach, the proposed method improved the evaluation values of the solution. The authors believe that their network-oriented modeling and solution approach is promising for developing a practical computer system for the rescheduling problem that would effectively support train recovery operations under strict time limitations. E.28 (Hollis, 2011), Ch. 4 related papers: (Hollis et al., 2006)54 practical problem type of transportation problem mode of transportation number of depots objectives size/practicality of the problem degree of urbanization regularity of the timetable admission of changeovers distribution road traffic multiple depot cost reduction application-based a range of postal and courier organizations, primarily Australia Post maximum of 1,016 trips (23 depots, 3 vehicle types55) urban/ex-urban both metropolitan (urban) and national level (ex-urban) distribution networks are considered irregular timetable with a cyclical nature; the network is usually repeated (with minor operational variations) each weekday restricted changeovers drivers are only allowed to drive vehicles based at their home depot56 solution method degree of integration degree of network segmentation complete integration whole network at once the largest instances tackled can be seen as a whole network model set covering The formulation used is based on set covering for duties, with an embedded circulation for vehicles. algorithm The author uses restricted enumeration followed by column generation to solve the linear relaxation of the proposed formulation. The enumeration heuristic is used to quickly provide a good set of initial columns. Column generation, incorporating connection fixing, is then used to solve the linear relaxation. Connection fixing, or branching on follow-ons, during column generation (this combination is sometimes referred to as price-and-branch57) is a heuristic fixing rule which involves identifying specific pairs of tasks and forcing them to appear consecutively in a driver duty. A dynamic programming algorithm is proposed to solve the associated column generation subproblem, generating specifically timed driver 54 Although (Hollis et al., 2006) mainly describes the combination of routing and scheduling, the scheduling problem in itself is also discussed. 55 Maximum of 8 vehicle types in another instance with less trips. 56 The computational study performed in (Hollis et al., 2006) investigated the effect of relaxing this restriction (i.e. relaxation to the case of unrestricted changeovers) for Australia Post product delivery networks showing that, on average, it resulted in an improvement of only 0.08%. Clearly this negligible improvement does not outweigh: the loss of operational simplicity, the exponential increase in the number of variables required when this restriction is relaxed, and a potential increase in algorithmic complexity for the column generation pricing problem. (Hollis et al., 2006) 57 Price-and-branch is not to be confused with branch-and-price where column generation is incorporated within branch-and-bound to prove optimality. E.29 duties. This algorithm is capable of: efficiently handling starting time windows for tasks, modeling generic driver duty regulations when the tasks covered have starting time windows (where task starting times can slide freely within their associated starting time window), and incorporating the added complexity of a concurrent vehicle scheduling problem. Finally, an integer solution is found from the set of duties produced by the enumeration and column generation process using branch-and-bound. dynamism of the solution approach static approach conclusion The results show that, on average, solutions where tasks can slide freely within their associated starting time windows improve upon those where task starting times are fixed (i.e. equivalent to a public transport context) by 7%. Furthermore, the former approach results in an average improvement of 2% and is over six times faster than the time window discretisation approach (used in (Hollis et al., 2006)). Very high quality solutions (within 0.05% on average of the best solution found) can be generated on average three times faster, by employing only an intelligent subset of the domination criteria (used for the elimination of labels in the dynamic programming algorithm). E.30 Appendix F (Hollis et al., 2006) “describe a simultaneous vehicle and crew routing and scheduling application for urban letter mail distribution at Australia Post. They are the first to consider a problem with multiple depots, where vehicles and drivers may be stationed and interchanged. The authors use a two-stage approach. In the first stage, they determine ‘abstract’ vehicle routes by solving a pickup-and-delivery problem with time windows, multiple depots, a heterogeneous fleet of vehicles as well as several working time restrictions for drivers. A pathbased mixed-integer programming (MIP) model is presented and solved by heuristic column generation. In the second stage, concrete vehicle and crew schedules are determined taking an integrated vehicle and crew scheduling approach. This is again done by solving an MIP with heuristic column generation. In the second-stage MIP, the tasks to be performed correspond to the vehicle routes computed in the first stage.” (Drexl et al., 2011, p. 4) “The authors examine the affect of different types of vehicle routing solutions on the vehicle and crew scheduling solution, comparing the different levels of integration (with respect to the set of allowable depots for a vehicle schedule covering a particular vehicle route) that are possible with the new vehicle and crew scheduling algorithm.” (Hollis et al., 2006, p. 133) “The solution technique employed […] has been shown to find high quality solutions. […] On average, simultaneous solutions improve upon those found using the sequential method by 8.31%. […] The algorithms and solution techniques presented in the paper have been used by network planners at Australia Post to demonstrate a potential transport network operational cost saving of 10% for the 2003 Melbourne metropolitan mail distribution network. Australia Post is currently using software developed based upon the ideas presented in this paper to assist in the management of ongoing changes to the mail distribution networks in major cities throughout Australia.” (Hollis et al., 2006, p. 149) (Xiang et al., 2006) “describe a static dial-a-ride problem which involves the scheduling of heterogeneous vehicles and a group of drivers with different qualifications. The solution procedure is a heuristic composed of a construction phase to obtain an initial solution, an improvement phase, and an intensification phase to fine-tune the solution. The important aspect of the procedure is that, initially, ‘abstract’ routes with a fixed schedule are determined, and only in the last stage, concrete vehicles and drivers are assigned to the routes.” (Drexl et al., 2011, p. 4) “The performance of the heuristic was evaluated by intensive computational tests on some randomly generated instances, […] revealing that this method is capable of quickly obtaining F.1 acceptable results. […] More specific, small gaps to the lower bounds from the column generation method were obtained in very short time for instances with no more than 200 requests. Although the result is not sensitive to the initial solution, the computational time can be greatly reduced if some effort is spent to construct a good initial solution. With this good initial solution, larger instances up to 2,000 requests were solved in less than 10 hours on a popular personal computer.” (Xiang et al., 2006, p. 1117) “Moreover, the method is flexible to cope with many practical constraints and does not contain any case-sensitive empirical parameter.” (Xiang et al., 2006, p. 1136) (Laurent & Hao, 2007) “consider the problem of simultaneously scheduling vehicles and drivers for a limousine rental company. The required transports are pickup-and-delivery trips with given time windows. The authors use a two-stage solution approach which aims to find a feasible crew and vehicle schedule by assigning a driver-limousine pair to each trip. First, an initial feasible solution is constructed by means of a greedy heuristic similar to the well-known best-fitdecreasing strategy for the bin packing problem, using constraint programming techniques for domain reduction. Second, an improvement procedure based on local search embedded in a simulated annealing metaheuristic is performed.” (Drexl et al., 2011, p. 3) “Results obtained on real data sets show a significant improvement in terms of quality, operational costs and elaboration time, compared with the actual practice in the examined company. Within a short time, the software supplies very good quality schedules in which the major part of the trips is assigned, satisfying all constraints. The approach also proves to be flexible. It unifies the treatment of the static and dynamic parts of this problem in a single framework. The decision support system based on this research is operating in the company and proves to be extremely useful. […] Although this paper deals with a real-world driver and vehicle scheduling problem in a particular application context where some aspects are specific, others are general ones. Therefore, the proposed method can be relevant to many other scheduling applications. For instance, the presented constraint-based model and some solution techniques of this work have been successfully transposed to the domain of transportation by bus in rural areas (see (Laurent & Hao, 2008) in section 2.1.3 and Appendix E).” (Laurent & Hao, 2007, p. 557) (Zäpfel & Bögl, 2008) “consider an application of local letter mail distribution. Pickup routes and delivery routes (but no combined pickup-and-delivery routes) have to be planned within a planning horizon of one week. In pickup routes, outbound shipments are transported from local post offices to a letter mail distribution centre. Conversely, in delivery routes, shipments are transported from the distribution centre to post offices. Schedules are planned for both drivers and vehicles, taking into account European Union social legislation. The problem is solved F.2 heuristically, by decomposing it into a Generalized VRP with Time Windows (GVRPTW) and a Personnel Assignment Problem (PAP). First, a feasible solution to the GVRPTW is computed […]. Then, the PAP tries to find a feasible driver assignment for the GVRPTW solution. The assignment is achieved by creating a table with all feasible combinations of drivers and routes. Each table entry represents the costs resulting from the driver performing the route. A complete personal assignment is computed, using three different strategies, among them a greedy and a random procedure. After that, an improvement procedure embedded into a metaheuristic follows.” (Drexl et al., 2011, p. 4) “Computer simulations have demonstrated that the constructed heuristic is suitable for practice.” (Zäpfel & Bögl, 2008, p. 980) “It is shown that especially an embedded Tabu Search procedure is very competitive for solving real problems as considered in the paper. […] All in all, this Tabu Search procedure can solve a complex decision problem with a tremendous number of variables and constraints in an efficient manner.” (Zäpfel & Bögl, 2008, p. 995) (Kim et al., 2010) “study a combined vehicle routing and staff scheduling problem where a certain number of tasks has to be fulfilled in a fixed sequence at customers. Among the tasks, an end-to-start relationship is assumed. In order to fulfill the tasks, different teams of workers are available. Each team is qualified to perform one specific type of task. The teams cannot move by themselves; instead, a set of vehicles is used to transport the teams. There is no fixed assignment of a vehicle to a team, and each vehicle may carry at most one team at a time. The authors develop an astonishingly simple procedure, in which the vehicles, the teams, and the next tasks for each customer are stored in three lists, along with the relevant information on times and locations. In each iteration, a triplet (vehicle, team, task) is selected from the lists, using a best-fit criterion. Then, the lists are updated to reflect the situation resulting when the selected vehicle transports the selected team to the location of the selected task.” (Drexl et al., 2011, p. 4) No relevant computational results were presented. “Good lower bounds for the problem still need to be developed to evaluate the solution quality of the proposed algorithm. […] Although, it is shown that the proposed approach significantly outperforms a simple greedy algorithm […] and that the solution quality is not too sensitive to the random numbers that it uses.” (Kim et al., 2010, pp. 8429-8430) (Prescott-Gagnon et al., 2010) “study the problem of planning oil deliveries to customers by lorry. To solve the problem, the authors develop three metaheuristics, a Tabu Search (TS) algorithm, a Large Neighborhood Search (LNS) heuristic based on this TS algorithm and another LNS heuristic based on a Column Generation (CG) heuristic which uses the TS algorithm to generate columns. They use a greedy construction heuristic which sequentially F.3 builds up routes for driver-lorry pairs by inserting the temporally closest customer. In the destruction phase of the LNS algorithm, different heuristics […] are used to determine the vertices to be removed from the routes of the current solution. The reconstruction phase applies either TS or CG. Among other move types, the TS procedure employs a driver switch move which tries to switch a pair of drivers, that is, have one driver drive the other driver's route and vice versa.” (Drexl et al., 2011, pp. 3-4) “Computational results obtained on instances derived from a real dataset indicate that the LNS methods outperform the TS heuristic. Furthermore, the LNS method based on CG tends to produce better quality results than the TS-based LNS heuristic, especially when sufficient computational time is available. However, the LNS-TS heuristic can be considered as a good alternative for a company that does not want to invest into a commercial linear programming solver that is required for the LNS-CG method.” (Prescott-Gagnon et al., 2010, p. 14) (Drexl et al., 2011) “study a simultaneous vehicle and crew routing and scheduling problem arising in long-distance road transport: pickup-and-delivery requests have to be fulfilled over a multi-period planning horizon by a heterogeneous fleet of lorries and drivers. They allow lorry/driver changes at geographically dispersed relay stations. […] European driver rules are considered completely and correctly for a planning horizon spanning one week. […] The solution approach is based on a heuristic decomposition of the overall problem into two stages, similar to the approach of (Hollis et al., 2006). […] In the first stage, routes for concrete lorries (as opposed to the abstract routes in (Hollis et al., 2006)) are determined, taking into account some driver rules. In the second stage, routes for drivers are computed, based on the lorry routes from the first stage and taking into account the remaining driver rules to ensure feasibility. […] Both stages can be solved with essentially the same rather simple and straightforward algorithm. This algorithm primarily relies on an appropriate network representation of the problem that is then solved using a large neighborhood search heuristic.” (Drexl et al., 2011, p. 1-6) “Extensive computational experiments have been performed with real-world data provided by a major freight forwarder. […] First and foremost, the presented algorithm is shown to be capable of solving large real-world instances and is able to achieve consistent and practically relevant results. […] However, for the given data set, lorry/driver changes offer no savings potential, so a fixed lorry-driver assignment seems to be the right set-up. These results, though, are only valid for the considered business field. The developed algorithm may well lead to very different results with other data or in other application areas.” (Drexl et al., 2011, p. 16-18) F.4 Nederlandse samenvatting In deze thesis behandelden we het routeren en plannen van voertuigen en personeel, enkele van de belangrijkste uitdagingen in hedendaagse transportbedrijven. Er werd dus getracht om het volledige planningsproces in de transportsector te beslaan. Vooraleer van start te gaan met de eigenlijke uiteenzetting, kozen we het jaar 2004 (meer precies 1 januari 2004) als het overgangspunt tussen verleden en heden. Deze keuze volgde vooral uit het feit dat de centrale onderzoeksvraag van deze thesis, ‘Levert het gebruik van een geïntegreerde benadering voor planningsproblemen in de transportsector (altijd) voordeel op?’, niet beschouwd werd vóór 2004. Voor de beschrijving van het verleden definieerden we eerst de individuele routering- en planningsproblemen, zijnde het voertuigrouteringprobleem (VRP), het voertuigplanningsprobleem (VSP) en het personeelsplanningsprobleem (CSP). Kortweg kunnen we stellen dat het VRP routes opbouwt zodat een aantal klanten bediend kunnen worden met een voertuigvloot, terwijl het VSP voertuigen aanduidt om die routes af te leggen en het CSP personeel toewijst aan de voertuigen (en de routes). De hoofddoelstelling is steeds het minimaliseren van de totale kosten, rekening houdend met bepaalde randvoorwaarden. Daarna situeerden we deze individuele problemen binnen het grotere geheel van het planningsproces in transportbedrijven. Dit planningsproces is opgedeeld in een strategische, een tactische en een operationele fase. Het routeren is een hoofdzakelijk strategisch aspect, daar waar het plannen (VSP en CSP) eerder deel uitmaakt van de operationele fase. We gingen ook verder in op het onderscheid tussen openbaar vervoer (transporteren van passagiers) en distributiesector (producten leveren aan klanten), net zoals het onderscheid tussen problemen met een enkel depot en met meerdere depots een belangrijke leidraad doorheen de thesis. In het openbaar vervoer zijn routes en dienstregelingen bijna steeds gegeven (bv. opgelegd door een lokale overheid) en blijven deze ongewijzigd gedurende lange periodes, wat niet het geval is in de distributiesector. Dit leidde tot het inzicht dat de combinatie van routering en planning over het algemeen een bezorgdheid is voor dit laatste type transportbedrijven, terwijl voor openbaar vervoer de focus vooral ligt op de planning van voertuigen en personeel. Het totale planningsproces in de distributiesector omvat dus meer aspecten, maar is intrinsiek zeker niet moeilijker of belangrijker dan zijn tegenhanger voor openbaar vervoer, waar de afweging tussen klanten en kosten veel delicater is. Het planningsprobleem werd beschreven, beginnend met de traditionele sequentiële aanpak (die in wezen neerkomt op het maken van de personeelsplanning, omdat dit altijd het eerst SV.1 opstellen van de voertuigplanning met zich meebrengt) en dan overgaand naar de geïntegreerde voertuig- en personeelsplanning (VCSP), inclusief een beknopt overzicht van de literatuur van vóór 2004. Het VCSP werd gedefinieerd in beide contexten – openbaar vervoer en distributie – om zo de gelijkenissen en verschillen aan te kunnen duiden. Een belangrijk aspect is de mogelijkheid om rekening te houden met tijdsvensters in een distributiecontext, wat niet van toepassing is in openbaar vervoer waar klanten een exact tijdstip van vertrek en aankomst verwachten. Een hoger niveau van integratie wordt nagestreefd wanneer we de routering en de planning simultaan proberen aan te pakken, we spreken dan over het voertuig- en personeels- routeringen planningsprobleem (VCRSP). Aangezien het routeringaspect ook deel uitmaakt van dit probleem, treedt het enkel op in de distributiesector. Het werd aangetoond dat een VCRSP gezien kan worden als een soort fusie van een VRP en een VCSP. De VCSP papers uit het heden werden gecategoriseerd aan de hand van een nieuwe procedure voorgesteld in deze thesis. Uiteraard dienden we eerst de categorisatiecriteria te definiëren – die ofwel betrekking hebben op het beschouwde praktisch probleem, ofwel op de oplossingsmethode die ervoor gebruikt wordt. In volgorde van dalende belangrijkheid, identificeerden we 8 criteria van de eerste soort (type transportprobleem, transportmiddel, aantal depots, doelstellingen, grootte/bruikbaarheid van het probleem, verstedelijkingsgraad, regelmatigheid van de dienstregeling en toelating van overstappen) en 5 van de tweede (graad van integratie, graad van netwerksegmentatie, model, algoritme en dynamisme van de oplossingsmethode). De eigenlijke categorisatie werd dan uitgevoerd voor 18 van de belangrijkste VCSP benaderingen, samen met de vermelding van een tiental papers die er sterk mee samenhangen. We kozen ervoor om enkel VCSP benaderingen voor wegverkeer – de meest besproken vervoerswijze – te beschouwen, aangevuld met één paper over spoorverkeer, maar geen luchtvaart. Uitbreiding van de categorisatieprocedure voor luchtverkeer en ook voor spoorverkeer is dus een evident werkpunt. Enkele verbanden tussen de voorgestelde criteria werden ontdekt en toegelicht: · bij stedelijke scenario’s hebben we meestal te maken met een enkel depot en bij regionale scenario’s met meerdere depots, · de doelstelling van dienstverleningsniveau (het beperken van vertragingen) kan verwezenlijkt worden met behulp van een dynamische planningsaanpak, en · de kwaliteit (regelmatigheid) van personeelsschema’s is enkel een relevante doelstelling als de dienstregeling onregelmatig is. SV.2 Ook werden een aantal interessante trends geobserveerd voor het VCSP, als resultaat van de categorisatie: · de overgrote meerderheid van auteurs passen de door hen voorgestelde methodes toe op bestaande levensechte problemen, · de grootste problemen tot nu toe beschouwd zijn van een grootte van 1414 ritten en deze groottes zijn de laatste jaren niet echt toegenomen terwijl dit voor het aantal depots en voertuigtypes wel het geval was, · er worden duidelijk meer regionale scenario’s beschouwd, · zo goed als alle methodes maken gebruik van complete integratie (tegenover partiële integratie), en · een set partitioning formulering is nog steeds het populairst en oplossingsmethodes gebaseerd op kolomgeneratie blijven het vaakst gebruikt waarbij een evolutie kan worden ontdekt in de richting van lineaire relaxatie (tegenover Lagrange relaxatie) voor de oplossing van het hoofdprobleem en branch-and-bound methodes (tegenover Lagrange heuristieken) om bruikbare oplossingen te bekomen. Uiteindelijk toonden we ook aan dat men een duidelijk en onbetwistbaar antwoord kan geven op het deel van de centrale onderzoeksvraag dat gerelateerd is aan het VCSP: ‘Levert het gebruik van een geïntegreerde benadering voor het voertuig- en personeelsplanningsprobleem (altijd) voordeel op?’. Dit antwoord is een volmondige ‘Ja’. Voor de beschrijving van de recente ontwikkelingen van het VCRSP verstrekten we een overzicht van de bestaande benaderingen (7 relevante papers) en zetten we een eerste stap naar een uitgebreide categorisatie gelijkaardig aan die voor het VCSP. Een tweedimensionale classificatie werd geïntroduceerd waarbij de dimensies overeenstemmen met de toelating van overstappen en de graad van integratie, waaruit opnieuw de twee overkoepelende categorisatie-aspecten praktisch probleem en oplossingsmethode blijken. Ook al is de bestaande literatuur over het VCRSP misschien nog niet uitgebreid genoeg opdat een doorgedreven categorisatie van nut zou zijn op dit moment, toch is het zeker een interessante opportuniteit voor de toekomst. Er werden een paar betekenisvolle observaties gedaan omtrent het VCRSP, namelijk dat de concrete toepassingsgebieden zeer gevarieerd zijn (limousineverhuur, postbedeling, oliedistributie, om er maar enkele te noemen) en dat nagenoeg elke publicatie tot dusver een partieel geïntegreerde aanpak gebruikt (let op het contrast met het VCSP). Het aan het VCRSP gerelateerde deel van de centrale onderzoeksvraag, ‘Levert het gebruik van een geïntegreerde benadering voor het voertuig- en personeels- routering- en planningsprobleem (altijd) voordeel SV.3 op?’, kan niet zo vastberaden beantwoord worden als het deel dat betrekking heeft tot het VCSP. We kunnen niet onomwonden ‘Ja’ antwoorden, maar nuanceren dit door te stellen dat er zeker potentieel zit in het integreren van routering en planning, op het eerste zicht minder dan voor het VCSP, maar in elk geval is er meer onderzoek nodig omtrent het VCRSP zodat we tot een universele conclusie kunnen komen. Dit verdere onderzoek kan daarenboven leiden tot nieuwe methodes die mogelijk wel steeds betere resultaten opleveren dan de traditionele benadering, zodat we dan wel volmondig ‘Ja’ kunnen antwoorden op de centrale onderzoeksvraag, net zoals bij het VCSP. Het viel buiten het opzet van deze thesis om verbeteringen voor te stellen voor specifieke bestaande oplossingsmethodes, meer bepaald voor modellen en algoritmes. Liever identificeerden we meer algemene en probleeminherente onderwerpen voor toekomstig onderzoek. Deze werden in de eerste plaats ontdekt door het identificeren van leemtes binnen de VCSP categorisatie en de tweedimensionale VCRSP classificatie en dan aanbevelingen te doen voor het opvullen van die leemtes. Wanneer voorhanden, vermeldden we ook enkele citaten betreffende elk onderzoeksonderwerp, letterlijk onttrokken aan de beschouwde papers teneinde de relevantie van de onderwerpen te staven. Enkele van de belangrijkste aanbevelingen voor het VCSP zijn: · streven naar het oplossen van het gehele netwerk ineens voor elk transportbedrijf, · een kleine aandachtsverschuiving van klanttevredenheid (doelstelling van dienstverleningsniveau) naar welzijn van werknemers (doelstelling van kwaliteitsvol personeelsschema), · rekening houden met de wens naar meer regelmatige personeelsschema’s wanneer men te maken heeft met een onregelmatige dienstregeling, · de grenzen betreffende de grootte van de beschouwde problemen blijven verleggen, · meerdere personeelstypes beschouwen in plaats van standaard slechts één enkel personeelstype te onderstellen, · het mogelijke voordeel van het toelaten van onbeperkt overstappen van crew onderzoeken (tegenover beperkte overstappen), · verder onderzoek uitvoeren in het gebied van metaheuristieken en constraint programming, en · meer aandacht besteden aan de ontwikkeling van dynamische planningsmethodes (en meer specifiek, een dergelijke methode introduceren in de distributiecontext). Het is opmerkelijk dat vele auteurs het belang van het laatst vernoemde onderzoeksonderwerk lijken te erkennen, hoewel we vaststelden dat slechts zeer weinig papers ook daadwerkelijk SV.4 zo’n dynamische aanpak hanteren. We identificeerden hier dus een hoge nood en een lage beschikbaarheid, met als gevolg dat dynamische planning zeker gezien mag worden als een heet hangijzer. We kwamen ook tot de conclusie dat het nodig zou kunnen blijken om wat extra aandacht te laten uitgaan naar het VCRSP in zijn geheel, aangezien het onderontwikkeld is ten opzichte van het VCSP, terwijl distributieproblemen niettemin een essentieel deel uitmaken van de hedendaagse transportactiviteiten. Andere significante aanbevelingen voor het VCRSP zijn: · ontwikkelen en evalueren van methodes die complete integratie nastreven, · testen van bestaande theoretische methodes op praktische problemen, · meer focussen op problemen met meerdere depots waar mogelijke overstaplocaties niet beperkt zijn tot één bepaalde plaats, en · VCRSP methodes aanpassen die initieel ontworpen waren voor een specifiek toepassingsgebied zodat ze meer algemeen toepasbaar worden en ook andere praktische problemen kunnen behandelen. Dit laatste onderzoeksonderwerp werd ook het meest aangehaald in andere papers. Wat uiteraard niet verbazend is, aangezien het een zeer algemene aanbeveling betreft die in feite de logische levenscyclus van een nieuw ontwikkelde oplossingsmethode weerspiegelt. SV.5