Chapter 6 TABU SEARCH HEURISTICS FOR THE VEHICLE ROUTING PROBLEM

Jean-Frangois Cordeau and Gilbert Laporte

Abstract

This article reviews some of the most important tabu search heuristics for the vehicle routing problem. Some of the main tabu search features are first described: neighbourhood structures, short term memory, long term memory, intensification. The tabu search algorithms are then described, followed by computational results and the conclusion. Vehicle Routing Problem, Tabu Search, Heuristics

Keywords:

1.

Introduction

The classical Vehicle Routing Problem (VRP) is defined on an undirected graph G = (V,E) where V = {vo,vi,... ,vn} is a vertex set and E = {(vi,Vj) : V{,Vj E V, i < j} is an edge set. Vertex vo represents a depot at which are based m identical vehicles of capacity Q, while the remaining vertices are cities or customers. A non-negative cost, distance or travel time matrix C = (cij) is defined on E. Each customer V{ has a non-negative demand qi and a non-negative service time s^. The VRP consists of determining a set of m vehicle routes i) of minimum total cost; ii) starting and ending at the depot; and such that iii) each customer is visited exactly once by exactly one vehicle; iv) the total demand of any route does not exceed Q; v) the total duration of any route does not exceed a preset bound D. The number of vehicles is either an input value or a decision variable. In the latter case vehicle fixed costs are sometimes incorporated in the objective function. Most authors in the field work with a fixed number of vehicles. It should be noted, however, that the best

146 solutions reported (in terms of distance) do not always use the least possible number of vehicles. The VRP is an important combinatorial optimization problem. It also occupies a central place in distribution management. Toth and Vigo (2002) report that the use of computerized methods in distribution processes often results in savings ranging from 5% to 20% in transportation costs. Golden, Assad and Wasil (2002) and Barker (2002) describe several case studies where the application of VRP algorithms has led to substantial cost savings. The VRP was introduced by Dantzig and Ramser (1959) more than four decades ago. Since then there has been a steady evolution in the design of solution methodologies, both exact and approximate, for this problem. Yet, no known exact algorithm is capable of consistently solving to optimality instances involving more than 50 customers (Golden et al., 1998; Naddef and Rinaldi, 2002). Heuristics are usually used in practice. Heuristics include constructive heuristics (e.g., Clarke and Wright, 1964) which gradually build a feasible solution while attempting to keep solution cost as low as possible, two-phase heuristics in which customers are first clustered into feasible routes and actual routes are then constructed (e.g., Gillett and Miller, 1974; Fisher and Jaikumar, 1981), and improvement methods which either act on single routes by the application of a Traveling Salesman Problem (TSP) heuristic, or on several routes by performing customer reallocations or exchanges. The best known intra-route moves are the classical r-opt interchanges (Croes, 1958; Lin, 1965) and several variants such as Or-opt (Or, 1976), 2-opt* (Potvin and Rousseau, 1995), and 4-opt* (Renaud, Boctor and Laporte, 1996). Inter-route exchanges include those of Fahrion and Wrede (1990), Thompson and Psaraftis (1992), Van Breedam (1994), and Kinderwater and Savelsbergh (1997). As reported by Laporte and Semet (2002), classical heuristics such as the savings method (Clarke and Wright, 1964) and the sweep algorithm (Gillett and Miller, 1974) usually have a high execution speed but often produce solutions having a large gap with respect to the best known solution value (typically between 3% and 7%). In the last fifteen years, several metaheuristics have been put forward for the solution of the VRP. These typically perform a thorough exploration of the solution space, allowing deteriorating and even infeasible intermediate solutions. A number of methods maintain a pool of good solutions which are recombined to produce even better ones. Gendreau, Laporte and Potvin (2002) have identified six families of metaheuristics for the VRP: simulated annealing, deterministic annealing, tabu search, genetic algorithms, ant systems, and neural networks. While the success of any particular method is related to its implementation features, it is fair to say that tabu search (TS) has been highly successful when applied to the VRP. For comparative computational results obtained by means of classical and metaheuristics over the Christofides, Mingozzi

Tabu Search Heuristics for the Vehicle Routing Problem

147

and Toth (1979) (CMT) fourteen benchmark instances, see Gendreau, Laporte and Potvin (2002), and Laporte and Semet (2002). Our purpose is to compare and analyze some of the best TS implementations for the VRP. We will first describe in Section 2 four important features of TS algorithms. This will be followed in Section 3 by an analysis of the individual TS heuristics with respect to these features and by an assessment and conclusion in Section 4.

2.

Tabu search features

Tabu Search was proposed by Glover (1986) and has quickly become one of the best and most widespread local search methods for combinatorial optimization. It performs an exploration of the solution space by moving from a solution xt identified at iteration t to the best solution xt+\ in a subset of the neighbourhood N(xt) of xt. Since xt+i does not necessarily improve upon xt, a tabu mechanism is put in place to prevent the process from cycling over a sequence of solutions. A nave way to prevent cycles is to forbid the process from going back to previously encountered solutions, but doing so would typically require excessive bookkeeping. Instead, some attributes of past solutions are registered and any solution possessing these attributes may not be considered for 6 iterations. This mechanism is often referred to as short term memory. Other features such as diversification and intensification are often implemented. The purpose of diversification is to ensure that the search process will not be restricted to a limited portion of the solution space. A possible way to implement diversification is to keep track of past solutions and penalize frequently performed moves. This mechanism was first applied to the VRP by Taillard (1992), and is often called long term memory. Intensification consists of performing an accentuated search around the best known solutions. Several survey papers and books have been written on TS, among which Hertz and de Werra (1991), Glover and Laguna (1993, 1997), and Gendreau (2003). In what follows, we examine neighbourhood structures, short term and long term memory strategies, and intensification methods in the VRP context.

2.1

Neighbourhood structures

There are two main ways to define neighbourhood structures in TS algorithms for the classical VRP. The first, termed A-interchanges by Osman (1991, 1993) consists of exchanging up to A customers between two routes. The second, called ejection chains (Rego and Roucairol, 1996) typically acts on more than two routes at the same time. In A-interchanges, the value of A is often restricted to 1 or 2 in order to limit the number of possibilities. Such operations are best described by couples (\, A2) (with Ai < A and A2 < A) to represent an operation where Ai customers are

148 moved from route 1 to route 2, and A2 customers are moved from route 2 to route 1. Thus, disregarding symmetries, the following moves are possible in 2exchanges: (2,2), (2,1), (2,0), (1,1), (1,0). Note that A-interchanges include swaps (1,1) and shifts (1,0). Several variants of this scheme are possible, such as limiting the set of routes or vertices considered for a A-interchange, or performing a local reoptimization of the routes in which removals or insertions take place. Some of the classical improvement procedures described in Section 1 could be applied in this context. Note that A-interchanges do not encompass intra-route moves but could be generalized to allow for this possibility. Ejection chains (Xu and Kelly, 1996; RegoandRoucairol, 1996; Rego, 1998) consist of first identifying a set of routes and then moving vertices in a cyclic manner from route k\ to route &2, from route &2toroute 3, etc. Because the fc^s are not necessarily different, ejection chains in fact allow for both intra-route and inter-route moves. Designing efficient TS algorithms entails striking a good balance between the quality of the neighbourhood and computational efficiency. More complicated neighbourhoods yield an enriched search since more solutions are considered at each iteration, but at the same time there is a risk that the computational effort will become excessive. This can be circumvented if sparse data structures are used, as in the granular tabu search algorithm (Toth and Vigo, 2003), or if the level of an ejection chain is limited, as in Rego and Roucairol (1996) or Rego (1998) where the complexity of each step is reduced from O(r?) to O(n2). Another issue related to neighbourhoods is whether intermediate infeasible solutions are considered during the search process. One advantage of allowing such moves is that they enable sequences of the type (xt, 2^+1, xt+2) where x% and xt+2 are both feasible, but xt+\ is infeasible and xt+2 is better than xt. One way of handling infeasibilities, proposed by Gendreau, Hertz and Laporte (1994), is through the use of a penalized objective function of the type

c{x) = c(x) + aQ(x)

where c(x) is the routing cost of solution x, Q{x) and D(x) measure the total excess demand and total excess duration of all routes, respectively, and a and (3 are self-adjusting parameters. Initially set equal to 1, these parameters are periodically reduced (increased) if the last fi solutions were all feasible (all infeasible), and // is a user-controlled parameter. The advantage of considering infeasible moves is greater in contexts where simple neighbourhood structures are used, such as (1,0) moves. When sophisticated neighbourhood schemes are employed, like ejection chains, the search process may be able to proceed from xt to xt+2 directly, without going through the infeasible solution

Tabu Search Heuristics for the Vehicle Routing Problem

149

2.2

Short term memory

To prevent cycling, it is common to prohibit reverse moves for 9 iterations. Thus a customer moved from route r to route s at iteration t may be prohibited from being reinserted in route r until iteration t + 9, or it may not leave route s until iteration t + 9. The first documented application of this now common rule can be found in Osman (1991). In Osman (1991, 1993) a fixed value of 9 is determined through regression analysis. In his study on the Quadratic Assignment Problem Taillard (1991) suggested varying the value of 9 during the search. This idea was applied to the VRP by Taillard (1992, 1993) and Gendreau, Hertz and Laporte (1994) who suggest selecting, at each iteration, the value of 9 in a given interval [0,9] according to a uniform distribution. Other authors, e.g., Cordeau, Laporte and Mercier (2001) use the same value of 9 throughout the search. It is also common to use an aspiration criterion to revoke the tabu status of a move if this causes no risk of cycling, for example if this yields a better overall incumbent feasible solution, or a better incumbent among the set of solutions possessing a certain attribute.

2.3

Long term memory

To diversify the search, most TS implementations penalize frequently used solution attributes or frequently performed moves (Glover, 1989). This can be done by adding to the routing cost c(xt+\) of xt+i a penalty term equal to the product of three factors: 1) a factor measuring the past frequency of the move; 2) a factor measuring instance size (such as \fn); 3) a user-controlled scaling factor. In practice, implementing such a mechanism is both effective and computationally inexpensive. This idea was first proposed by Glover (1989) and later fine tuned by Taillard (1992, 1993, 1994) and by Gendreau, Hertz and Laporte (1994).

2.4

Intensification

Intensification consists of accentuating the search in promising regions. Periodic route improvements by means of TSP algorithms may be classified as intensification techniques. Another way of performing intensification is to conduct a search using a wider neighbourhood structure around some of the best known solutions. The concept of Adaptive Memory is probably the most powerful intensification tool for TS. An adaptive memory is a population of good solutions encountered during the search procedure. It was applied to the VRP for the first time by Rochat and Taillard (1995). Similar to what is done in genetic algorithms, these authors combine solution elements from the population to

150 construct new solutions. If their value is less than those of some solutions in the population, they are retained and the worst solutions are discarded from the population. In the VRP context, a new solution is initialized from a number of non-overlapping routes extracted from good solutions. Typically these routes do not cover all vertices for otherwise there would be overlaps. The search is then initiated from the selected routes and the unrouted vertices.

3.

An overview of TS algorithms

We now present an overview of some of the most popular TS algorithms for the VRP, with an emphasis on the most innovative features of each algorithm.

3.1

Osman's algorithm

It was Osman (1991, 1993) who proposed the concept of A-interchanges. In his implementation he uses A = 1 to generate an initial feasible solution and A = 2 in the descent phase. The value A = 2 allows a mix of single and double vertex moves, and single and double vertex swaps between vehicle routes. Osman tested two strategies for selecting a neighbour solution. In the first, called best admissible (BA), the best non-tabu solution is selected. In the second, called first best admissible (FBA), the first admissible improving solution is selected if one exists, otherwise the best admissible solution is retained. Osman shows through empirical testing that with the same stopping criterion FBA produces slightly better solutions, but this variant is much slower than BA. Osman's TS implementation uses fixed tabu tenures, no long-term memory mechanism and no intensification schemes.

3.2

The Gendreau, Hertz and Laporte Taburoute algorithm

In Taburoute (Gendreau, Hertz and Laporte, 1994) neighbour solutions are obtained by moving a vertex from its current route r to another route s containing one of its closest neighbours. Insertion into route s is performed concurrently with a local reoptimization, using the GENI mechanism for the TSP (Gendreau, Hertz and Laporte, 1992). This may result in creating a new route or in deleting one. To limit the neighbourhood size, only a randomly selected subset of vertices are considered for reinsertion in other routes. The concept of a penalized objective d(x) was introduced in Taburoute. The parameters a and (3 are adjusted every 10 iterations: if the ten previous solutions were feasible with respect to capacity (route duration), then a(/3) is divided by 2; if they were all infeasible, then a{(3) is multiplied by 2; otherwise, a((3) remains unchanged. Any vertex removed from route r at iteration t is prevented from being reinserted in that route until iteration t + 0, where 6 is randomly selected in [5,10], unless

Tabu Search Heuristics for the Vehicle Routing Problem

151

of course such a move would yield a new incumbent solution. A continuous diversification scheme is also applied. Routes are periodically updated using the US post-optimization heuristic for the TSP (Gendreau, Hertz and Laporte, 1992). Two types of intensification features are used. False starts are employed to perform a limited search from several starting solutions, and then running the main search procedure starting from the best solution obtained. In addition, after a given number of iterations without improvement, a more intense search is performed starting with the incumbent, but allowing this time more vertices to be moved from their current route.

3.3

Taillard's algorithm

While it was published one year earlier then Taburoute, Taillard's (1993) algorithm was developed during the same period. It uses a 1-interchange mechanism without local reoptimization, and without allowing infeasibilities. This scheme is much simpler than that used in Taburoute and allows for more iterations to be performed during the same running time. The tabu mechanism and the long-term memory policy are the same as in Taburoute. Periodically, routes are reoptimized using the exact TSP algorithm of Volgenant and Jonker (1983). The search procedure developed by Taillard employs a decomposition scheme that lends itself to the use of parallel computing. In planar problem, the set of customers is first partitioned into sectors centered at the depot, and also into concentric circles. Search is performed in each subregion by a different processor. The subregion boundaries are updated periodically to provide a diversification effect. In non-planar problems, regions are defined through the computation of shortest spanning arborescences rooted at the depot. Taillard's algorithm remains to this day one of the TS heuristics for the VRP yielding the best solution values. On the fourteen CMT instances, it has produced twelve of the best known solutions.

3.4

The Rochat and Taillard Adaptive Memory Procedure

The Adaptive Memory Procedure (AMP) implemented by Rochat and Taillard was presented under the title "Probabilistic Diversification and Intensification". The method should not be regarded as a VRP heuristic per se, but rather as a general procedure applicable to several contexts and in conjunction with several heuristic schemes. For example, it was applied by Bozkaya, Erkut and Laporte (2003) to postoptimize political districts obtained by means of a TS heuristic. Its application to the VRP has helped improve two of the solutions generated by Taillard on the CMT instances, thus yielding the best known incumbent for thirteen of the fourteen instances.

152

3.5

The Xu and Kelly algorithm

In their TS algorithm, Xu and Kelly (1996) define neighbours by oscillating between ejection chains and vertex swaps between two routes. The ejection chains are determined by minimizing a flow on a network of the type depicted in Figure 6.1. Level 2 and level 3 arcs have a flow of 1 or 0 to indicate whether a customer is removed from a route (level 2) or reinserted in a route (level 3).
Customer modes Route nodes

Level 1 Figure 6.1.

Level 2

Level 3

Level 4

Network flow problem for the Xu and Kelly algorithm.

Since several customers can be removed from the same route or reinserted into the same route, all arc costs are approximations only. As in some of the previous algorithms, individual routes are periodically reoptimized by means of 3-opt and 2-opt operations. Intermediate infeasible solutions with respect to capacity are allowed (with a = 1), fixed tabu tenures are used, and a long-term frequency based memory is applied. A pool of best known solutions is maintained and the search procedure is periodically applied to the solutions of this pool in the hope of generating even better ones. Overall, this algorithm has produced several good solutions on the capacity constrained VRP but is rather time consuming and complicated because it requires the setting of several parameters.

3.6

The Rego and Roucairol Tabuchain algorithm

The Rego and Roucairol Tabuchain algorithm (1996) uses ejection chains involving levels, or routes, to define neighbourhoods. Basically an ejection process bumps a vertex from one level to the successive level, starting at level 1 and ending at level 1. At any level of the ejection chain, the last bumped

Tabu Search Heuristics for the Vehicle Routing Problem

153

vertex may possibly be relocated in the position of the vertex removed from level 1, but also possibly elsewhere (Figure 6.2).

Figure 6.2. An ^-level ejection chain. Three vertices are shown for each route. The middle one is ejected.

An ejection chain is considered only if no arc (edge) appears more than once in the solution, but routes that violate capacity or duration constraints are accepted. A sequential and a parallel version of this algorithm were implemented by the authors.

3.7

Rego's Flower algorithm

The Flower algorithm described by Rego (1998) applies an ejection process to move from a VRP solution made up of several blossoms (or routes) to another solution by suitably deleting and introducing edges. Applying these operations create intermediate structures consisting of one path, called stem, emanating from the depot, and several blossoms attached to it. Thus a blossom may be transformed into a stem (Figure 6.3a) or divided into a blossom and a stem (Figure 6.3b). During this process, the number of active vehicle routes may vary. Candidate solutions consisting only of blossoms are obtained by performing ejection moves that maintain the flower structures. This is done by suitably deleting edges from the structure and inserting new ones. This process is embedded within a TS mechanism by declaring tabu the reintroduction of an edge that has just been deleted. As in some of the previous algorithms, variable tabu tenures are used and a long-term diversification strategy based on move frequencies is applied. Here, the penalty term is randomly generated in an interval whose length is related

3.8

The Toth and Vigo granular tabu search algorithm

The granularity concept proposed by Toth and Vigo (2003) does not only apply to the VRP or to TS algorithms, but to discrete optimization on the whole. Like AMP, it is a highly portable mechanism. The idea is to permanently remove from consideration long edges that have only a small likelihood of belonging

154

a) Transforming a blossom into a stem by removing an edge incident to the depot.

b) Transforming a blossom into a blossom and a stem by deleting an edge and introducing another one.

Figure 6.3. Two ejection chain rules in the Flower algorithm. In both cases the depot is represented by the black vertex.

to an optimal solution. More specifically, a threshold is defined and the search is performed on the restricted edge set E(u) = {(^, Vj) G E : Cij < u} U / , where / is a set of important edges defined as those incident to the depot. The value of v is set equal to pc9 where f3 is called a sparsification parameter, and c is the average edge cost in a good feasible solution quickly obtained, for example, by the Clarke and Wright (1964) algorithm. In practice, selecting (3 in the interval [1.0,2.0] results in the elimination of 80% to 90% of all edges. Toth and Vigo have applied this idea in conjunction with some of the main concepts proposed by Taillard (1993) and by Gendreau, Hertz and Laporte (1994).

3.9

The Unified Tabu Search Algorithm of Cordeau, Gendreau, Laporte and Mercier

The Unified Tabu Search Algorithm (UTSA) was first developed in the context of the Periodic VRP and of the Multi-Depot VRP (Cordeau, Gendreau and Laporte, 1997) for which it produced highly competitive results. It was later applied to the Site Dependent VRP with time windows (Cordeau and Laporte, 2001), to the Periodic VRP with and the Multi-Depot VRP with time windows (Cordeau, Laporte and Mercier, 2001) and, finally, to the classical VRP with capacity and route length constraints (Cordeau et al., 2002). UTSA uses several of the features of Taburoute, namely the same neighbourhood structure, GENI insertions, long-term frequency based penalties. It also differs from Taburoute in a number of ways. The search is applied to a single initial solution, fixed tabu durations are used and no intensification phase is applied. In other words

Tabu Search Heuristics for the Vehicle Routing Problem

155

an efficiency gain was achieved by sacrificing some of the sophistication of Taburoute. While UTSA also allows intermediate infeasible solutions, it either divides or multiplies a and /? by a factor 1 + 5 at each iteration according to whether the previous solution was feasible or not with respect to either capacity or route duration. The tabu mechanism operates on an attribute set B(x) = {(i, k) : i\ is visited by vehicle k in solution x}. Neighbour solutions are obtained by removing (i, k) from B(x) and replacing it with (i, A/), where k' ^ k. Attribute (i, k) is then declared tabu for a number of iterations, and the aspiration criterion is defined relative to attribute (i, k). Two interesting features of UTSA are its relative simplicity (it operates with very few parameters) and its flexibility. It has been successfully applied to a large variety of VRP variants by always keeping its parameters at the same value.

4.

Assessment and conclusion

Tabu search clearly stands out as one of the best heuristics for the VRP. Over the last fifteen years several implementations have been developed and tested. It is fair to say they have been highly successful in tackling this difficult problem. This success is due in part to a number of key ideas contained in several implementations: efficient neighbourhood evaluation through proper data structures, the allowance of infeasible solutions during the search, the use of self-adjusting parameters, continuous diversification, ejection chains, adaptive memory, and granularity. More than any other feature, these ideas have contributed to the success of TS heuristics and have withstood the test of time. We present in Table 6.1 comparative computational results for the implementations described in Section 3, over the fourteen CMT benchmark instances. While a proven optimum is known for the first instance only, the best known solutions obtained by Rochat and Taillard (1995) and by Mester and Bnysy (2005) are believed to be near-optimal and are used to assess the degree of accuracy of other heuristics. Most TS implementations used in the comparison yield average deviations of less than 1% from the best known. Computation times are difficult to compare since they are not always reported, different computers are used, some implementations use parallel computing (Taillard, 1993; Rego and Roucairol, 1996), and it is not always clear how many runs were made. For example, Taillard (1993) and Rochat and Taillard (1995) only report the time required to reach a solution whose value is within a given percentage of the best known. In the case of parallel implementations, an upper bound on the running time of an equivalent sequential algorithm can be found by adding up the time taken by each processor. It is worth mentioning that the recent

156 study by Dongarra (2004) can help compare, to some extent, the relative speed of computers. We believe there is little scope for major accuracy improvements in the future for the CMT instances. Instead, some effort should be devoted to developing leaner implementations, using fewer parameters and simpler design, even if accuracy deteriorates slightly as a result. Also, algorithms should be easily adaptable to handle the multiplicity of side constraints that arise in practical contexts. As noted by Cordeau et al. (2002) the research community is used to assessing heuristics with respect to accuracy and speed alone, but simplicity and flexibility are often neglected. Yet these two attributes are essential to adoption by end-users and should be central to the design of the next generation of algorithms.

Acknowledgements
This research was partly funded by the Canadian Natural Sciences and Engineering Research Council (NSERC) under grants 227837-00 and OGP0039682, and by the Fonds pour la Formation de chercheurs et l'aide a la recherche (FCAR) under grant 2002-ER-73080. This support is gratefully acknowledged. Thanks are due to Cesar Rego and to a referee for their valuable comments.

156

8*

Table 6.1. Summary of computational results

I3

3.
33

Osman Value2 Deviation Minuted 524.61 0.00 1.90 844.00 0.84 1.05 1.44 25.72 838.00 1044.35 59.33 1.55 1334.55 3.35 54.10 2.88 555.43 0.00 911.00 0.15 17.61 1.39 878.00 49.99 1184.00 76.26 1.85 76.02 1441.00 3.23 24.07 1043.00 0.09 14.87 819.59 0.00 1547.00 47.23 0.38 0.00 866.37 19.60

Taburoute Value 4 Deviation Minuted 6.00 0.00 524.61 53.80 0.06 835.77 0.40 18.40 829.45 0.75 58.80 1036.16 2.43 90.90 1322.65 0.00 555.43 13.50 2.37 54.60 931.23 865.94 0.00 25.60 1.31 71.00 1177.76 1.62 99.50 1418.51 1073.47 3.01 22.20 819.56 0.00 16.00 0.44 1547.93 59.20 0.00 65.70 866.37

Taillard Rochat and Taillard Value Deviation Minutes7 Value Deviation Minutes? 0.18 0.82 524.61 0.00 0.00 524.61 1.13 0.00 0.88 835.26 0.00 835.26 15.00 9.67 826.14 0.00 826.14 0.00 30.00 63.33 1028.42 0.00 0.00 1028.42 0.58 0.01 50.00 1291.45 1298.79 _ 0.00 0.28 555.43 0.00 555.43 0.00 0.85 909.68 0.00 909.68 0.00 18.33 865.94 0.00 865.94 0.00 18.33 1162.55 0.00 1162.55 1395.85 0.15 0.00 1397.94 1042.11 45.00 0.00 0.00 1042.11 5.83 5.67 819.56 0.00 0.00 819.56 _ 0.00 65.00 1541.14 0.00 1541.14 25.00 866.37 0.00 0.00 866.37

Acknowledgements
16.19 16.19

3-

Instance n Type1 1 50 C C 2 75 3 100 c 4 150 c 5 199 c 6 50 C,D 7 75 C,D 8 100 C,D 9 150 C,D 10 199 C,D C 11 120 C 12 100 13 120 CD 14 100 CD Average % deviation from best and time for all instances for C instances for C,D instances 33,60 25.83 41.37 0.89 0.95 0.82 46.80 38.01 55.59 0.05 0.08 0.02 21.51 21.73 21.30 0.00 0.00 0.00

1.03 1.07 1.00

This research was partly funded by the Canadian Natural Sciences and Engineering Research Council (NSERC) under grants 227837-00 and OGP0039682, and by the Fonds pour la Formation de chercheurs et l'aide a la recherche (FCAR) under grant 2002-ER-73080. This support is gratefully acknowledged. Thanks are due to Cesar Rego and to a referee for their valuable comments.

study by Dongarra (2004) can help compare, to some extent, the relative speed of computers. We believe there is little scope for major accuracy improvements in the future for the CMT instances. Instead, some effort should be devoted to developing leaner implementations, using fewer parameters and simpler design, even if accuracy deteriorates slightly as a result. Also, algorithms should be easily adaptable to handle the multiplicity of side constraints that arise in practical contexts. As noted by Cordeau et al. (2002) the research community is used to assessing heuristics with respect to accuracy and speed alone, but simplicity and flexibility are often neglected. Yet these two attributes are essential to adoption by end-users and should be central to the design of the next generation of algorithms.

1. 2. 3. 4. 5. 6. 7.

C: Capacity restrictions; D: Duration restrictions. FBA version. Minutes on a VAX 8600 computer. Standard algorithm with one set of parameters. Silicon Graphics Workstation (36 MHz, 57 MFlops). Silicon Graphics 4D/35. Time required to reach a solution within 1% of the best known. Silicon Graphics Indigo (100 MHz). Time required to reach a solution within 1% of the best known.

156

Summary of computational results Xu and Kell> Tabuchain (sequential) Value Deviation Minuted* Value Deviation Minutes3 524.61 0.00 4.89 524.61 0.00 1.05 835.26 0.00 37.94 837.50 0.01 43.38 826.14 47.45 827.53 0.17 36.72 0.00 0.11 100.81 1054.29 1.55 48.47 1029.56 3.35 77.07 1298.58 0.56 207.80 1338.49 555.43 0.00 2.38 909.68 0.00 82.95 868.29 0.09 18.93 1178.84 0.14 29.85 1420.84 1.79 42.72 1042.1.1 2.80 1043.54 0.00 0.00 11.23 819.56 1.57 0.00 5.61 819.56 0.00 1550.17 0.59 1.95 _ 866.53 0.00 24.65 Tabuchain (parallel) Value Deviation Minuted 524.61 0.00 0.85 835.32 0.27 16.80 0.17 33.90 827.53 2.52 27.20 1044.35 3.66 16.25 1334.55 555.43 0.00 3.17 23.10 909.68 0.00 866.75 0.27 8.60 1164.12 1.40 15.55 1420.84 1.79 52.02 0.14 6.30 1042.11 819.56 0.00 1.22 1550.17 0.59 2.00 0.02 9.42 866.37 0.10 58.19 0.77 0.96 0.58 30.21 31.36 29.06 0.55 0.72 0.37 15.46 14.65 16.27

Acknowledgements

Table 6.1.

(continued).

Instance n Type 1 50 C 2 75 C 3 100 C 4 150 C 5 199 C 6 50 C,D 7 75 C,D 8 100 C,D 9 150 C,D 10 199 C,D 11 120 C 12 100 C 13 120 C,D 14 100 C,D Average % deviation from best and time for all instances for C instances for C,D instances

Flower Value Deviation Minutesw 524.81 0.04 0.17 847.00 1.41 0.38 832.04 0.93 0.71 1047.21 1.83 1.91 4.64 1351.18 7.06 559.25 0.69 0.06 922.21 1.38 0.50 876.97 1.27 1.41 1191.30 2.47 6.07 1460.83 4.66 5.06 1052.04 0.95 3.96 821.63 0.25 0.54 1558.06 3.69 1.10 867.79 0.75 0.16

1.54 1.40 1.68

2.32 2.14 2.51

This research was partly funded by the Canadian Natural Sciences and Engineering Research Council (NSERC) under grants 227837-00 and OGP0039682, and by the Fonds pour la Formation de chercheurs et l'aide a la recherche (FCAR) under grant 2002-ER-73080. This support is gratefully acknowledged. Thanks are due to Cesar Rego and to a referee for their valuable comments. study by Dongarra (2004) can help compare, to some extent, the relative speed of computers. We believe there is little scope for major accuracy improvements in the future for the CMT instances. Instead, some effort should be devoted to developing leaner implementations, using fewer parameters and simpler design, even if accuracy deteriorates slightly as a result. Also, algorithms should be easily adaptable to handle the multiplicity of side constraints that arise in practical contexts. As noted by Cordeau et al. (2002) the research community is used to assessing heuristics with respect to accuracy and speed alone, but simplicity and flexibility are often neglected. Yet these two attributes are essential to adoption by end-users and should be central to the design of the next generation of algorithms.

8. DEC Alpha workstation (DEC OSF/1 v3.0). 9. Sun Sparc IPC. The value reported for the parallel version is the total time on four of these computers. 10. HP 9000/712.

156

Table 6.1. Best 105

(continued).

3.

Acknowledgements

Summary of computational results Granular Tabu Search UTSA Value Deviation Minutes1'2Value Deviation Minutes1 i 524.61 0.00 0.81 524.61 0.00 4.57 838.60 2.21 835.45 0.02 7.27 0.40 828.56 0.29 0.40 11.23 2.39 829.44 1033.21 18.72 0.47 4.51 1038.44 0.97 1318.25 2.09 7.50 1305.87 1.13 28.10 555.43 0.00 0.86 555.43 0.00 4.61 920.72 1.21 2.75 909.68 0.00 7.55 869.48 0.41 11.17 2.90 866.38 0.05 1173.12 0.91 5.67 1171.81 19.17 0.80 1435.74 9.11 1415.40 1.40 29.74 2.86 1042.87 0.07 S.I 8 1074.13 14.15 3.07 819.56 0.00 0.00 10.99 1.10 819.56 1545.51 0.28 1.80 14.53 9.34 1568.91 866.37 0.00 1.41 866.53 0.02 10.65 524.61 835.26 826.14 1028.42 1291.29 555.43 909.68 865.94 1162.55 1395.85 1042.11 819.56 1541.14 866.37 3.84 3.10 4.58 0.69 0.80 0.58 13.75 13.58 13.92

Instance n Type 1 50 C 2 75 C 3 100 C 4 150 C 5 199 C 6 50 C,D 7 75 C,D 8 100 C,D 9 150 C,D 10 199 C,D 11 120 C 12 100 C 13 120 C,D 14 100 C,D Average % deviation from best and time for all instances for C instances for C,D instances

0.64 0.47 0.81

This research was partly funded by the Canadian Natural Sciences and Engineering Research Council (NSERC) under grants 227837-00 and OGP0039682, and by the Fonds pour la Formation de chercheurs et l'aide a la recherche (FCAR) under grant 2002-ER-73080. This support is gratefully acknowledged. Thanks are due to Cesar Rego and to a referee for their valuable comments. study by Dongarra (2004) can help compare, to some extent, the relative speed of computers. We believe there is little scope for major accuracy improvements in the future for the CMT instances. Instead, some effort should be devoted to developing leaner implementations, using fewer parameters and simpler design, even if accuracy deteriorates slightly as a result. Also, algorithms should be easily adaptable to handle the multiplicity of side constraints that arise in practical contexts. As noted by Cordeau et al. (2002) the research community is used to assessing heuristics with respect to accuracy and speed alone, but simplicity and flexibility are often neglected. Yet these two attributes are essential to adoption by end-users and should be central to the design of the next generation of algorithms.

11. Pentium (200 MHz). 12. Sun Ultrasparc 10 (440 MHz). 13. References in which the best known solution values were first reported: 1, 6, 14 (Osman, 1991); 2, 3, 4, 7, 8, 9, 11, 12, 13 (Taillard, 1993); 5 (Mester and Biaysy, 2005); 10 ( Rochat and Taillard, 1995). Some best known solution values may have been previously published in working papers, theses, etc. The solution value of instance 1 is known to be optimal (Hadjiconstantinou, Christofides andMingozzi, 1995).

160

References
Barker, E.K. (2002) "Evolution of Microcomputer-Based Vehicle Routing Software: Case Studies in United States," in The Vehicle Routing Problem, P. Toth and D. Vigo (eds), SIAM Monographs on Discrete Mathematics and Applications, Philadelphia, 353-362. Bozkaya, B., E. Erkut and G. Laporte (2003) "A Tabu Search Heuristic and Adaptive Memory Procedure for Political Districting," European Journal of Operational Research, 144:12-26. Christofides, N., A. Mingozzi and P. Toth (1979) "The Vehicle Routing Problem," in Combinatorial Optimization, N. Christofides, A. Mingozzi and P. Toth (eds), Wiley, Chichester, 315-338. Clarke, G. and J.W. Wright (1964) "Scheduling of Vehicles from a Central Depot to a Number of Delivery Points," Operations Research, 12:568-581. Cordeau, J.-F., M. Gendreau and G. Laporte (1997) "A Tabu Search Heuristic for Periodic and Multi-Depot Vehicle Routing Problems," Networks, 30:105119. Cordeau, J.-F., M. Gendreau, G. Laporte, J.-Y. Potvin and F. Semet (2002) "A Guide to Vehicle Routing Heuristics," Journal of the Operational Research Society, 53:512-522. Cordeau, J.-F. and G. Laporte (2001) "A Tabu Search Algorithm for the Site Dependent Vehicle Routing Problem with Time Windows," INFOR, 39:292298. Cordeau, J.-F., G. Laporte and A. Mercier (2001) "A Unified Tabu Search Heuristic for Vehicle Routing Problems with Time Windows," Journal of the Operational Research Society, 52:928-936. Croes, G.A. (1958) "A Method for Solving Traveling-Salesman Problems," Operations Research, 6:791-812. Dantzig, G.B. and J.H. Ramser (1959) "The Truck Dispatching Problem," Management Science, 6:80-91. Dongarra, J.J. (2004) "Performance of Various Computers Using Standard Linear Equations Software," Technical Report CS-89-85. Computer Science Department, University of Tennessee. Fahrion, R. and M. Wrede (1990) "On a Principle of Chain-Exchange for Vehicle-Routeing Problems (1-VRP)," Journal of the Operational Research Society, 41:821-827. Fisher, M.L. and R. Jaikumar (1981) "A Generalized Assignment Heuristic for the Vehicle Routing Problem," Networks, 11:109-124. Gendreau, M. (2003) "An Introduction to Tabu Search," in Handbook ofMetaheuristics, F. Glover and G.A. Kochenberger (eds), Kluwer, Boston, 37-54.

Tabu Search Heuristics for the Vehicle Routing Problem

161

Gendreau, M., A. Hertz and G. Laporte (1992) "New Insertion and Postoptimization Procedures for the Traveling Salesman Problem," Operations Research, 40:1086-1094. Gendreau, M., A. Hertz and G. Laporte (1994) "A Tabu Search Heuristic for the Vehicle Routing Problem," Management Science, 40:1276-1290. Gendreau, M., G. Laporte and J.-Y. Potvin (2002) "Metaheuristics for the VRP," in The Vehicle Routing Problem, P. Toth and D. Vigo (eds), SIAM Monographs on Discrete Mathematics and Applications, Philadelphia, 129-154. Gillett, B.E. and L.R. Miller (1974) "A Heuristic Algorithm for the Vehicle Dispatch Problem," Operations Research, 22:340-349. Glover, F. (1986) "Future Paths for Integer Programming and Links to Artificial Intelligence," Computers & Operations Research, 13:533549. Glover, F. (1989) "Tabu Search - Part I," ORSA Journal on Computing, 1:190206. Glover, F. and M. Laguna (1993) "Tabu Search," in Modern Heuristic Techniques for Combinatorial Problems, C.R. Reeves (ed.), Blackwell, Oxford, 70-150. Glover, F. and M. Laguna (1997) Tabu Search. Kluwer, Boston. Golden, B.L., A.A. Assad and E.A. Wasil (2002) "Routing Vehicles in the Real World: Applications in the Solid Waste, Beverage, Food, Dairy, and Newspaper Industries," in The Vehicle Routing Problem, P. Toth and D. Vigo (eds), SIAM Monographs on Discrete Mathematics and Applications, Philadelphia, 245-286. Golden, B.L., E.A. Wasil, J.P. Kelly and I-M. Chao (1998) "Metaheuristics in Vehicle Routing," in Fleet Management and Logictics, T.G. Crainic and G. Laporte (eds), Kluwer, Boston, 33-56. Hadjiconstantinou, E.A.,N. Christofides and A. Mingozzi (1995) "A New Exact Algorithm for the Vehicle Routing Problem Based on g-Paths and fc-Shortest Paths Relaxation," Annals of Operations Research, 61:21-43. Hertz, A. and D. de Werra (1991) "The Tabu Search Metaheuristic: How We Used It," Annals of Mathematics and Artificial Intelligence, 1:111-121. Kinderwater, G.A.P. and M.W.P. Savelsbergh (1997) "Vehicle Routing: Handling Edge Exchanges," in Local Search in Combinatorial Optimization, E.H.L. Aarts and J.K. Lenstra (eds). Wiley, Chichester, 337-360. Laporte, G. and F. Semet (2002) "Classical Heuristics for the Capacitated VRP," in The Vehicle Routing Problem, P. Toth and D. Vigo (eds), SIAM Monographs on Discrete Mathematics and Applications, Philadelphia, 109-128. Lin, S. (1965) "Computer Solution of the Traveling Salesman Problem," Bell System Technical Journal, ^llAS-llff}. Mester, D. and O. Braysy (2005) "Active Guided Evolution Strategies for the Large Scale Vehicle Routing Problems with Time Windows," Computers & Operations Research,forthcoming.

162 Naddef, D. and G. Rinaldi (2002) "Branch-and-Cut Algorithms for the VRP," in The Vehicle Routing Problem, P. Toth and D. Vigo (eds), SIAM Monographs on Discrete Mathematics and Applications, Philadelphia, 53-84. Or, I. (1976) "Traveling Salesman-Type Combinatorial Problems and their Relation to the Logistics of Regional Blood Banking," Ph.D. Thesis, Northwestern University, Evanston, IL. Osman, I.H. (1991) "Metastrategy Simulated Annealing and Tabu Search Algorithms for Combinatorial Optimization Problems," Ph.D. Thesis, The Management School, Imperial College, London. Osman, I.H. (1993) "Metastrategy Simulated Annealing and Tabu Search Algorithms for the Vehicle Routing Problem," Annals of Operations Research, 41:421-451. Potvin, J.-Y. and J.-M. Rousseau (1995) "An Exchange Heuristic for Routing Problems with Time Windows,"Journal ofthe Operational Research Society, 46:1433-3446. Rego, C. (1998) "A Subpath Ejection Method for the Vehicle Routing Problem," Management Science, 44:1447-1459. Rego, C. and C. Roucairol (1996) "A Parallel Tabu Search Algorithm Using Ejection Chains for the Vehicle Routing Problem," In Meta-Heuristics: Theory and Applications, Kluwer, Boston, 661-675. Renaud, J., F.F. Boctor and G. Laporte (1996) "A Fast Composite Heuristic for the Symmetric Traveling Salesman Problem," INFORMS Journal on Computing, 8:134-143. Rochat, Y. and E.D. Taillard (1995) "Probabilistic Diversification and Intensification in Local Search for Vehicle Routing," Journal of Heuristics, 1:147 167. Taillard, E.D. (1991) "Robust Taboo Search for the Quadratic Assignment Problem," Parallel Computing, 17:433^145. Taillard, E.D. (1992) "Parallel Heuristic Search Methods for the Vehicle Routing Problem," Working Paper ORWP 92/03, Departement de mathematiques, Ecole Federate Polytechnique de Lausanne, Switzerland. Taillard, E.D. (1993) "Parallel Iterative Search Methods for Vehicle Routing Problem," Networks, 23:661-673. Taillard, E.D. (1994) "Parallel Taboo Search Techniques for the Job Shop Scheduling Problem," ORSA Journal on Computing, 6:108-117. Thompson, P.M. and H.N. Psaraftis (1993) "Cyclic Transfer Algorithms for Multi-Vehicle Routing and Scheduling Problems," Operations Research, 41:935-946. Toth, P. and D. Vigo (2003) "The Granular Tabu Search and its Application to the Vehicle Routing Problems," INFORMS Journal on Computing, 15:333-346.

Tabu Search Heuristics for the Vehicle Routing Problem

163

Toth, P. and D. Vigo (2002) "An Overview of Vehicle Routing Problems," in The Vehicle Routing Problem, P. Toth and D. Vigo (eds), SIAM Monographs on Discrete Mathematics and Applications, Philadelphia, 1-26. Van Breedam, A. (1994) "An Analysis of the Behavior of Heuristics for the Vehicle Routing Problem for a Selection of Problems with Vehicle-Related, Customer-Related, and Time-Related Constraints," Ph.D. Dissertation, University of Antwerp, 1994. Volgenant, A. and R. Jonker (1983) "The Symmetric Traveling Salesman Problem and Edge Exchanges in Minimal 1-Tree," European Journal of Operational Research, 12:394-403. Xu, J. and J.P. Kelly (1996) "A Network Flow-Based Tabu Search Heuristic for the Vehicle Routing Problem," Transportation Science, 30:379-393.