Applying the Ant System to the Vehicle Routing Problem
Bernd Bullnheimer, Richard F. Hartl and Christine Strauss

Department of Management Science, University of Vienna Bruenner Str. 72, A-1210 Vienna, Austria Email(s): fbullnhei, hartl, straussg@pom.bwl.univie.ac.at

In this paper we use a recently proposed metaheuristic, the Ant System, to solve the Vehicle Routing Problem (VRP) in its basic form, i.e. with capacity and distance restrictions, one central depot and identical vehicles. A \hybrid" Ant System algorithm is rst presented and then improved using problem speci c information (savings, capacity utilization). Experiments on various aspects of the algorithm and computational results for fourteen benchmark problems are reported and compared to those of other metaheuristic approaches such as Tabu Search, Simulated Annealing and Neural Networks.

The Ant System, introduced by Colorni, Dorigo and Maniezzo 6], 10], 12] is a new distributed metaheuristic for hard combinatorial optimization problems and was rst used on the well known Traveling Salesman Problem (TSP). It has further been applied to the Job Shop Scheduling Problem in 7], to the Graph Colouring Problem in 8] and to the Quadratic Assignment Problem in 18]. Observations on real ants searching for food were the inspiration to imitate the behaviour of ant colonies for solving combinatorial optimization problems. Real ants are able to communicate information concerning food sources via an aromatic essence, called pheromone. They mark the path they walk on by laying down pheromone in a quantity that depends on the length of the path and the quality of the discovered food source. Other ants can observe the pheromone trail and are attracted to follow it. Thus, the path will be marked again and will therefore attract more ants. The pheromone trail on paths leading to rich food sources close to the nest will be more frequented and will therefore grow faster. The described behaviour of real ant colonies can be used to solve combinatorial optimization problems by simulation: arti cial ants searching the solution space simulate real ants searching their environment, the objective values correspond to the quality of the food sources and an adaptive memory corresponds to the pheromone trails. In addition, the arti cial ants are equiped with a local heuristic function to guide their search through the set of feasible solutions. In this paper we present the application of the ant system to the Vehicle Routing Problem (VRP) with one central depot and identical vehicles. The remainder of the paper is organized as follows: In x2 we present the VRP and the ant system algorithm to tackle it. A "hybrid" ant system algorithm, using the 2-opt heuristic and problem speci c information, is developed in x3 and x4, respectively. Experiments on various
Sophia-Antipolis, France, July 21-24, 1997 INRIA & PRiSM-Versailles

1 Introduction

g. the arti cial ants construct vehicle routes by successively choosing cities to visit. To solve the VRP. 19]). genetic algorithms (e. For the selection of a (not yet visited) city. The VRP and the TSP are closely related. v2. 1997 . ij = 1=dij . and how promising is the choice of that city. As soon as the customers of the VRP are assigned to vehicles. 14]) have been proposed to solve it. an information that is stored in the pheromone trails ij associated with each arc (vi . France. We conclude with a discussion of the results in x6.g. the VRP is reduced to several TSPs. The VRP is a very complicated combinatorial optimization problem that has been worked on since the late fties. v1 .g.MIC97 aspects of the algorithm and computational results for fourteen benchmark problems are presented in x5. The vertex v0 denotes the depot. The aim is to nd minimum cost vehicle routes where every customer is visited exactly once by exactly one vehicle all vehicle routes begin and end at the depot for every vehicle route the total demand does not exceed the vehicle capacity Q for every vehicle route the total route length (incl. the other vertices of V represent cities or customers. two aspects are taken into account: how good was the choice of that city. In the case of the VRP (or the TSP) it is de ned as the reciprocal of the distance. 17]) and neural networks (e. vj ) : i 6= j g is a set of arcs. vj ). For that reason our approach is highly in uenced by the TSP ant system algorithm by Dorigo et al. city vj is selected to be vistited after city vi according to a random-proportional rule 11] that can be stated as follows: 2 Ant System for VRPs 8 P ij ] ij ] > < ih ] pij = > h2 : 0 ih ] if vj 2 otherwise (1.2 2ND INTERNATIONAL CONFERENCE ON METAHEURISTICS . d) where V = fv0. 0 = 0).g. 12].g. service times) does not exceed a given bound L. July 21-24. 5]. the depot is chosen and a new tour is started. : : :. Whenever the choice of another city would lead to an infeasible solution for reasons of vehicle capacity or total route length. and the nonnegative weights dij . This latter measure of desirability. The Vehicle Routing Problem (VRP) can be represented by a complete weighted directed graph G = (V. until each city has been visited. is the local heuristic function mentioned above. Problem speci c methods (e. For each customer vi a nonnegative demand qi and a nonnegative service time i is given (q0 = 0. called visibility and denoted by ij . simulated annealing (e. which are associated with each arc (vi .e. vng is a set of vertices and A = f(vi .1) INRIA & PRiSM-Versailles Sophia-Antipolis. A. because of its central meaning in distribution management. 13]). With = fvj 2 V : vj is feasible to be visitedg fv0g. i. vj ). represent the distance (or the travel time or the travel cost) between vi and vj . 15]) as well as metaheuristics like tabu search (e.

Hybridization in general means combining ideas of two di erent methods in one approach. all arcs belonging to the so far best solution (objective value L ) are emphasized as if ants. respectively. In the next section the "hybrid" ant system is further improved by including problem speci c information in step II (a) of the algorithm.MIC97 3 This probability distribution is biased by the parameters and that determine the relative in uence of the trails and the visibility. so-called elitist ants had used them. Such proceeding is common practice for hard combinatorial optimization problems and has been successfully applied to other problems like e. the pheromone trail is increased by ij = 1=Lk . the two steps construction of vehicle routes and trail update. and zero otherwise. After initializing the basic ant system algorithm. Only then the total objective value is calculated and the trails are updated. The implication for the VRP is that as many ants are used as there are customers in the VRP (i. where m is the number of arti cial ants: new ij = Concerning the initial placement of the arti cial ants it was found that the number of ants should be equal to the number of cities in the TSP. For a more detailed description the reader is referred to 12]. m old X ij + k=1 k ij + ij (1.g. are repeated for a given number of iterations.2ND INTERNATIONAL CONFERENCE ON METAHEURISTICS . Furthermore.203. A tour is called 2-optimal if there is no possibility to shorten the tour by exchanging two arcs. each vehicle route generated is checked for 2-optimality and is improved if possible. One elitist ant increases the trail intensity by an amount ij that is equal to 1=L if arc (vi . p. and that each ant should start its tour from another city2. We use the term hybrid following Goldberg's schematic of a \hybrid using a batch scheme" 16]. Trail evaporation is used to avoid early convergence. where customers are rst clustered and then 2-optimal vehicle routes for each cluster are generated. France. For each arc (vi . In vehicle routing 2-opt is used in the SweepAlgorithm 15]. the construction of vehicle routes. 3 It is questionable whether the addition of the 2-opt approach deserves the name hybrid method or whether it is only a post-optimization. The same can be done with solutions constructed by arti cial ants: at the end of an ant system iteration. In addition to that. The 2-opt-heuristic for the TSP 9] is an exchange procedure that generates a socalled 2-optimal tour. July 21-24. 2 These are results of experiments by Dorigo et al. the pheromone trails are laid depending on the objective value Lk . 12] as well as our own experiments.2). the trail intensities are updated according to the following Formula (1. Sophia-Antipolis. vj ) belongs to the so far best solution. The "hybrid" ant system3 for the VRP can be described by the schematic algorithm given in Figure 3. timetabling 3] or production scheduling 20]. After an arti cial ant k has constructed a feasible solution.e. 1997 INRIA & PRiSM-Versailles . m = n). The resulting quantitative improvements achieved by the "hybrid" ant system are shown in x5. part of the existing pheromone trails evaporates ( is the trail persistence)1 . vj ) that was used by k ant k.2) 3 \Hybrid" Ant System Algorithm 1 Elitist ants improved the results obtained for the TSP and were therefore also used for the VRP.1. and that one ant is placed at each customer at the beginning of an iteration. Thus.

has been mentioned above. Let Qi the total capacity used including the capacity requirement of customer vi .10 ) This information is included in the visibility. the depot. In the VRP not only the relative location of two cities is important4 but also their relative location to the depot v0 is essential for the tour length. Then. July 21-24. Furthermore. that can be included in an ant system algorithm for the purpose of quality improvement of the solutions. High savings ij indicate that visiting customer vj after vi is a good choice. according to decreasing savings. m generate a new solution using Formula (1. it seems reasonable to assure a high degree of capacity utilization of the vehicles. The so-called savings5 measure the favourability of combining two cities vi and vj to a tour and can be quanti ed by: ij = di0 + d0j ? dij . namely the existence of one distinct city in the VRP. then high6 values ij = (Qi + qj )=Q indicate high capacity utilization through the visit of customer vj after visiting vi . if pij ij where the parameter regulates the relative in uence of the savings. This again can be used for the ant system by giving those customers a high probability of being selected: pij ij The parameter determines the relative in uence of ij . One major di erence.4 2ND INTERNATIONAL CONFERENCE ON METAHEURISTICS . and thus the corresponding ant system approaches. This can be used to improve the quality of the ant system algorithm if high savings lead to a high probability of selection.2) Figure 3. That approach starts with depot-customer-depot tours.e.1) (b) Improve all vehicle routes using the 2-opt-heuristic (c) Update the pheromone trails using Formula (1. tours are combined as long as no restrictions are violated. France. for a capacity restricted problem as the VRP.MIC97 I Initialize II For I max iterations do: (a) For each ant k = 1. i. : : :.1: "Hybrid" Ant System Algorithm 4 Problem Speci c Improvements The close relation between the VRP and the TSP. 1997 . The probability distribution for selecting customer vj to be visited next after customer vi can thus be extended to: 8 P ij ] > < pij = > h2 : ij ] ij ] ij ] ih ] ih ] ih ] ih ] if j 2 otherwise 0 (1. 6 ij 1 for a feasible solution. See the Savings-Algorithm in 5]. has been taken into account. 4 5 INRIA & PRiSM-Versailles Sophia-Antipolis. But the VRP has some further characteristics.

For the problems with bounded route length all customers require the same service time = 1 = : : : = n . France. which contains 50 randomly distributed customers. savings 7 For each experiment we simulated 30 independent ant system runs of 50 iterations each.MIC97 5 The ant system for VRPs was tested on fourteen benchmark problems described in 4]. In three tests we studied the sole in uence of respectively.32 12. whereas for the former it is not.85 3.52 7.42% 542. 1997 INRIA & PRiSM-Versailles .61 3.11 4. These problems contain between 50 and 199 customers in addition to the depot. The customers in problems C1-C10 are randomly distributed in the plane. The results showed clearly that using the trail information does contribute to the quality of the solution: without it the average objective value was 23% above the optimum. To see whether the pheromone trails contribute at all to the results we tested = 0. The same is true for the clustered problems: problems C13-C14 are the counterparts of problems C11-C12 with additional route length constraint. Problems C1-C5 and C6-C10 are identical. the average over 30 runs was 17% above the optimum. 9 The parameter setting = 1.44 7. we illustrate some of our experiments7 as well as the stepwise improvement of the ant system at hand of problem C1.59% 554. The "hybrid" ant system (HAS) on the other hand. As we used one ant per customer.01% 524. The best solution the ants found by selecting the customers according to the probability distribution given in Formula (1. 7%) than the basic ant system.31% 617.74 12.42 3. the number of solutions generated per iteration was equal to the number of customers.1.1) was 12%.18% 599.00% 5 Computational Results Table 5.88 1. Through the problem speci c features described in x4. through the use of Formula (1. generated much better solutions (dev. In the following there is no distinction made regarding this aspect.22 23. if not indicated otherwise. except that for the latter the total route length is bounded.58% 540.61 0. thus a total of 2500 solutions was generated per run. As a summary of the results we present the deviation from the optimal solution8 for the best and the average solution out of 30 runs in Table 5.2ND INTERNATIONAL CONFERENCE ON METAHEURISTICS . The basic ant system method NN AS HAS HAS-sav HAS-cap HAS-1 HAS-5 dev.61% 592. July 21-24.10) for the selection probabilities. a setting that could be described as a stochastic nearest neighbour heuristic (denoted by NN).43% 563. 646. i. while they are clustered in problems C11-C14.1) solved problem C1 just satisfactory in the rst experiment9 .48% 546.1: Comparison of Results algorithm (denoted by AS in Table 5.e.47 17.91% 564.67% 542. Footnote 2) as well as the VRP and was chosen. = 5 and = 0:75 lead to good results for the TSP (cf.10% 532.36 5.66 14.70% 590. a further reduction of route lengths was achieved. 8 Strictly speaking these solutions are not always the optimal but the "best published" solutions as only for some of them optimality has been proven. best dev. Before the results for these test problems are presented at the end of this section. Sophia-Antipolis.

The latter uses visibility.e. As a consequence of the reduced in uence of the pheromone trails compared to visibility. i.6% as compared to 7. no pheromone trails are used). HAS-1)10. the adaptive e ect almost vanished. The graph shows clearly that the local search & 2-opt procedure is outperformed by the "hybrid" ant system. In order to analyze the ant speci c contribution to the quality of the results. = 5.1: Ant System vs. HAScap). dev. Therefore all terms were weighted equally and the parameter setting = = = = 5 was chosen. dev. 4%). Both features improved the performance of the ant system algorithm. France.6 2ND INTERNATIONAL CONFERENCE ON METAHEURISTICS . savings and capacity utilization. Local Search Figure 5.1 depicts the continuous reduction of objective values (50 iterations = 2500 b solutions. Recall the similar comparison between the basic ant system (AS) and the stochastic nearest neighbour heuristic (NN). 1997 . we further compared the "hybrid" ant system (HAS-5) with a stochastic local search procedure11 . savings and capacity utilization. and worked best when applied simultaneously (avg.MIC97 ( = 5. = 0. and the 2-opt heuristic for tour improvement. the arti cial ants select the customers in this stage primarily according to visibility. 5.4%). 10 11 INRIA & PRiSM-Versailles Sophia-Antipolis. which is also done in the The other two parameters were kept at = 1 and = 5. Thus. averaged over 30 runs) for both methods. with the savings yielding better results (avg. as well as their combined in uence ( = = 5. July 21-24. = = = 5. 610 600 590 local search & 2-opt 580 HAS-5 570 560 550 540 1 2500 Figure 5. denoted by HAS-sav) and capacity utilization ( = 0.61). which lead to the best results where the ant system (HAS-5) found the optimal solution (total length 524. The trail intensities are still close to their initial value 0 and have therefore hardly any e ect on the selection probabilities. savings and capacity utilization for tour construction ( = 0. In the early phase of the search the two methods look almost identical.

5 compares the computational results for the fourteen test Sophia-Antipolis.38% Table 5. Thus.10% 0. Finally the in uence of the trail persistence was subject of further tests (cf.61 530. For lower values the learning e ect diminishes and even though the nding of very good solutions is possible.2: In uence of Elitist Ants In further tests we studied the in uence of the elitist ants. i. caused by massive exploration of suboptimal tours early in the search.42 545. initial placement dev. France. starting the search from there is another possibility because the depot is per de nition included in every vehicle route.2ND INTERNATIONAL CONFERENCE ON METAHEURISTICS .00% 527.39% Table 5.76 4. 5. Alternatively. Higher values prevent e cient exploration of the search space as the trail intensities on arcs belonging to suboptimal vehicle routes are kept too high for too long. In 12] the ant system performance for a TSP with 30 cities rst increased with the number of elitist ants (up to an optimal range around 8) and then decreased again. namely the depot. the ants use this accumulated information.73% 3.01% 3.4). Later.90 1. The following Table 5.61 0. 0 10 30 50 70 90 559.86% 3.42 3.00% random 545.70% 4. which is illustrated in Table 5. but only up to a range around 50.08% 1. the number of \regular" ants / customers. July 21-24.17 540. The results underline our early ndings that = 0:75 is a good setting. the average quality of the algorithm decreases. best dev. The comparison of these options. choosing the starting points for the arti cial ants randomly is also possible.84 5.13 524.26 531.98 0. Table 5.23% 0. when trail intensities for some arcs increase because of frequent use.12 544. the solution space is reduced and better solutions are generated.68% 0. The results for various numbers of elitist ants are illustrated in Table 5. and decrease for others because of evaporation.03% 531.74 550. 6.64% customer 540.94 dev.64% best 552.84 dev. As the VRP has one distinct city. 1997 INRIA & PRiSM-Versailles .3. con rms our assumption that placing one ant at each customer is best.96% 4.40 548. For the VRP we found a similar phenomenon: introducing elitist ants and increasing their number brought better results.3: In uence of Initial Placement Furthermore we looked at the initial placement of the arti cial ants.01% 524. whereas the local search is still based on initial data only.2.00% 1. depot 550.MIC97 7 local search procedure. The use of more elitist ants lead to poorer performance.20 525.04 528.e.

C11 C12 C13 C14 n 50 75 100 150 199 50 75 100 150 199 n 120 100 120 100 Q 160 140 200 200 200 160 140 200 200 200 L 200 160 230 200 200 1 1 1 1 1 Clustered problems optimal solution 1042.95 0.61 524. where the customers are clustered.91% 9 0.45% 8 1147.54% best 531. For each problem the columns give the problem size n.16 4.33 540.61 0. : : :.86 dev. For the problems C11-C14.68a 865.39% 14 1504.14a 1028.42 544.20% 12 0.40% 11 a b Taillard 24] Rochat and Taillard 22] Q L 200 1 0 200 1 0 200 720 50 200 1040 90 Table 5. For all problems I max = 100 iterations were simulated with = n elitist ants.43 6. The random problems were solved using HAS-5 ( = = = = 5). the vehicle capacity Q.57% 12 1473. the service time and the objective value of the optimal solution. vehicles used 2.MIC97 dev.52 869.02% 3. In the last three columns the best solutions obtained with the ant system.77% 4.56a 1541.80% 19 Ant System 1072.23% 12 886. vn. 1.23% 10 879.8 2ND INTERNATIONAL CONFERENCE ON METAHEURISTICS .66 525.42a 1291.14a 866.25 545.34% 0. the maximalroute length L.05% 10 3.00% 0.01 3.37a 0 0 0 0 0 10 10 10 10 10 optimal solution 524.76% 3.01% 3.41 548.26a 826. July 21-24. C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 Prob. According to our ndings we set = 0:75 and used m = n ants.68 544. France.96 1590.45b 555.63 524.58 4.09% 16 562.34% 9 1202.61a 835.85b Ant vehicles System dev.41 11.75 0.13 524.5: Ant System Results problems.93 dev.43a 909.99 0.94a 1162.06% 0.35% 6 948. the deviation from the optimum and the number of vehicles used are shown.00% 0. we found that the savings INRIA & PRiSM-Versailles Sophia-Antipolis.45 819.17 2.00% 5 870.93 1.10% 0.11a 819. used 524. initially placed at the customers v1 .79 7. 1997 .50 0.40 14.55a 1395. 4.42 Table 5.4: In uence of Trail Persistence Random problems Prob.

only for problems C4 and C5 the ant system showed higher deviations.4 7.30% with only 12 out of 14 problems tested.0 Sun Sparc 4 Clustered problems 0. July 21-24.6 C10 1.45% 11.79% 4.27% 8.3 2.9 C2 0. we used HAS-cap ( = 1.40% 3.73% 5.20%.37% 102.4 0.06% 4.4 2.3 0.35% 0. Tabu search (the sequential algorithm from 21]12) outperforms Tabu Prob.58% 40.00% 1.00% 5.1 C8 0. Thus. which are located behind each other.2 0.27% 16.2ND INTERNATIONAL CONFERENCE ON METAHEURISTICS . There the deviation from the optimum ranged from 0.68% 1.1 3.20% 4.2 0.09% but better than Ghaziri's neural networks approach 14] where the average deviation was 5.14% 6.59% 2.3 1.64% 16. Especially the results on the clustered problem instances C11-C14 seem to be better.5 6.02% 9.22% 52.47% 13.3 C6 0.80% 57.2 11.5 VAX 8600 5.6 5.14% 6.5 0.85% 4.78% 0.8 0.2 0.34% 7.00% 3.73% 27.00% 11.1 Pentium 100 4.7 4.9 8.2 3.23% 2.2 C5 3.79% 0.51% 23.2 | | 4.31% 76.6 1. C11 C12 C13 C14 Simulated Neural Ant Annealing 19] Networks 14] System 0.91% 16. Search 21] C1 0.05% to 3.55% 22.09% Neural Networks 14] 5.05% 10. The computational results show that reasonably good solutions can be obtained by the ant system.9 8.77%.2 14.0 0. 1997 INRIA & PRiSM-Versailles . Sophia-Antipolis.MIC97 9 do not really contribute to an improvement. = 5.6: Deviation and Run Times for several Metaheuristic Approaches all other metaheuristics with an average deviation of 0.23% 3.52% 27.77% VAX 8600 2. The reason is that cities belonging to di erent clusters.43%) performs not as good as Osman's simulated annealing approach 19] with 2.57% 28.09% 82.39% 26.55% 8.28% 18.9 0.9 C4 2.2 0. As run times are another criteria for the quality of an algorithm the proposed method is compared to other metaheuristic approaches for which run times were reported in Table 5. = 0 and = 5) for the clustered problems.2 C7 0.14% 983.79% 52.1 2.6 0.3 12.00% 0.3 1.6 C9 1.8 C3 0.5 6.30% Ant System 2. Most random problems were solved within a 5% range.4 2.00% 23.00% 0.6.1 3. The ant system (4.40% 15.4 0.88% 71.0 Prob.6 8.2 0.2 0.17% 33.6 1.37% 31. might be combined to a tour because of high savings (which result from being located in line with the depot).3 0.3 13.09% 6.3 Random problems Tabu Simulated Search 21] Annealing 19] 0. To ensure maximum comparability we did not include their parallel implementation.43% Table 5. France.65% 0.40% 59.4 | | 4. Run times (given in CPU minutes in 12 A comparison on basis of run times on di erent machines is not perfectly meaningful.

Also an automatic adjustment of the parameters like done e. demands.g. 23]. 6 Discussion and conclusion Acknowledgements The authors would like to thank Marco Dorigo.MIC97 Table 5. INRIA & PRiSM-Versailles Sophia-Antipolis. The presented contribution shows the application and the improvement of an ant system algorithm to the VRP. Primarily. low trail levels or both.g. a more detailed analysis of parameter values is necessary. Another very interesting aspect is the use of candidate lists. A more radical change of the existing algorithm would be to use the ants only to cluster the customers and subsequently. and. France. e. to apply a local search to nd good tours among them. July 21-24. even though our current version can not yet compete with the best tabu search algorithms. The superiority of tabu search for VRPs can be explained by two facts: tabu search is an excellent method that has been studied and improved a lot since its introduction. Tabu Search performs much better. but nevertheless the results for the ant system also indicate that there still is much potential for improvement. In addition to that. Besides these methodological considerations. As the ant system can compete with other vehicle routing metaheuristics in terms of run times.6) for all algorithms are more or less similar and vary with the problem size in a range from approximately one minute for the smallest to approximately one and a half hours for the largest problem. For many of them the probability of being selected is very low because of large distances. 22]. The computational results con rm the positive experiences made with the ant system by applying it to the TSP 1]. more elaborated local search procedures exchanging customers not only within but also among tours should be looked at. A metaheuristic could be used to guide the search through the parameter space. In the current version of the ant system all feasible customers have the chance to be selected. the presented approach is an alternative to tackle VRPs. A similar idea using a genetic algorithm as a cluster builder has been proposed in 17]. Moreover the algorithm seems to be well suited for parallel implementation 2]. multiple depots or problems with time windows are of interest. Although some very good solutions for the VRP instances were obtained. 11]. 1997 . 21]. Therefore we are certain that future work on the ant system approach will help to further improve its quality for vehicle routing. 19]. makes perfect planning impossible. Vittorio Maniezzo and Gerhard Waescher as well as two anonymous referees for their comments that helped to improve the quality of this paper.10 2ND INTERNATIONAL CONFERENCE ON METAHEURISTICS . service times etc. additional modi cations of the algorithm to extensions of the VRP. Concentrating only on the more promising candidates should yield better results. 24]) than on any other method. in Evolution Strategies might be of use for the ant system. For practical purposes deviations up to 5% are more than acceptable as uncertainty about travel costs. much more VRP related research has been done on tabu search (cf. the best known solutions for the fourteen test problems could not be improved. 13].

Morgan Kaufmann. Strauss. and L. Oper. A. Hertz. Toth. Scheduling of Vehicles from a Central Depot to a Number of Delivery Points. Gambardella. and P.Belgian Journal of Operations Research. A. 6-th Int. M. Kelly. Conf. and C. 46 June 1997. editors. and C. Sandi. Working Paper No. France. Maniezzo. Wright. Distributed Optimization by Ant Colonies. Ischia (Italy). editors. and C. and P. Laporte. A Hybrid Genetic Algorithm for Highly Constrained Timetabling Problems. pages 605-610. Dorigo. Colorni. In Proc. Bullnheimer. A new rank based version of the ant system: a computational study. Meta-Heuristics: Theory & Applications. pages 29-41. R. Dorigo. Management Sci. In F. 1 (1997) 1. Dorigo. and V. pages 39-53. 14] H. Optimization. pages 651-660. Kotsis. Elliman. Maniezzo. 13] M. Comput. 9] G. 10] M. Trubian. INRIA & PRiSM-Versailles Sophia-Antipolis. and A. J. 3] E. 6] A. Hertz. M. Politecnico di Milano. 1997. V. 11] M. Croes. Dorigo. pages 53-66. Combinatorial Optimization. 4] N. pages 791-812. Ant System: Optimization by a Colony of Cooperating Agents. Arti cial Life (ECAL'91). IEEE Trans.1. Conf. Mingozzi. Supervision in the Self-Organizing Feature Map: Application to the Vehicle Routing Problem. pages 1276-1290.2ND INTERNATIONAL CONFERENCE ON METAHEURISTICS . 5] G. Res. The Vehicle Routing Problem. Doctoral Dissertation. Oper. Wiley. 1992. G. 48 (1997). IEEE Trans. and J. Costa. and J. Weare. and R. Ants can colour graphs. JORBEL .F. 6 (1958). 1991. Colorni. Colorni. Paper presented at Conference on High Performance Software for Nonlinear Optimization: Status and Perspectives (HPSNO'97). 1995. July 21-24. Dorigo.W. SFB Adaptive Information Systems and Modelling in Economics and Management Science. Bullnheimer. Statistics and Computer Science 34 (1994) 1.. Mingozzi. Toth. Europ. 1996. Strauss. Christo des. Elsevier Publishing. Proc. D. pages 295-305. 12] M.K. Soc. P. 40 (1994). Maniezzo. A Tabu Search Heuristic for the Vehicle Routing Problem. Res. Oper. pages 315-338. Vienna. Man. Learning and Natural Algorithms. Burke. Clarke. and G. and M. Ant system for Job-Shop Scheduling. Christo des. pages 134-142. Hartl.G. 12 (1964). Res. pages 568-581. 1979. 8] D. Kluwer Academic Publishers. In N. Bourgine. Parallelization Strategies for the Ant System. Cybernetics 26 (1996) 1. Italy (in Italian). A. Ghaziri. editors. Gendreau. A Method for solving Traveling-Salesman Problems. In I. Sys. 2] B. 7] A. Ant Colony System: A Cooperative Learning Approach to the Traveling Salesman Problem.A. Genetic Algorithms (ICGA'95). Varela. and A. 1997 . Osman.F.MIC97 11 Bibliography 1] B. Evol.M. V.

Roucairol. Miller. editors. Parallel Iterative Search Methods for Vehicle Routing Problems. G. Taillard. July 21-24. Osman. Kelly. 24] E. pages 661-675. and E. France. Kluwer Academic Publishers. Heuristics 1 (1995). A. and E. MetaHeuristics: Theory & Applications. Optimization. 19] I.E. 1996. 1994. Res. The Ant System applied to the Quadratic Assignment Problem. The MAX-MIN Ant System and Local Search for the Traveling Salesman Problem. Addison-Wesley. and J. Belgium. J. Technical Report IRIDIA / 94-28.12 2ND INTERNATIONAL CONFERENCE ON METAHEURISTICS . INRIA & PRiSM-Versailles Sophia-Antipolis. Osman. and Machine Learning. 20] E. ICEC'97 . and H. 1997 . Dorigo. Kopfer. pages 308-313. Erkens. IEEE Press. A Heuristic Algorithm for the Vehicle Dispatch Problem. Networks 23 (1993). Entwicklung eines hybriden Genetischen Algorithmus zur Tourenplanug. Colorni. Goldberg. Physica. Oper. In I. 1994. 23] T. Maniezzo. Metastrategy simulated annealing and tabu search algorithms for the vehicle routing problem. Genetic Algorithms in Search. Gillett. Hoos. and L. and C. Taillard. 16 (1994). Proc.MIC97 15] B. Stuetzle. A Parallel Tabu Search Algorithm Using Ejection Chains for the Vehicle Routing Problem. pages 421-451.R. 21] C.1997 IEEE 4-th Int. Pesch. and M. Rochat. 17] H. 41 (1993). Res. 22 (1974) pages 340-347. Oper. Evolutionary Computation. Conf. Spekt. Ann. pages 661-673. Universite Libre de Bruxelles. 1989. Probabilistic Diversi cation and Intensi cation in Local Search for Vehicle Routing. Res. Pankratz. Learning in Automated Manufacturing. 22] Y. 16] D. 18] V. Oper. Rego. pages 147-167. pages 21-31.

Sign up to vote on this title
UsefulNot useful