Engineering Applications of Articial Intelligence 14 (2001) 825837

Articial intelligence heuristics in solving vehicle routing problems with time window constraints
K.C. Tana,*, L.H. Leeb, K. Oua
a

Department of Electrical and Computer Engineering, National University of Singapore, 4 Engineering Drive 3, 10 Kent Ridge Crescent, Singapore 119260 b Department of Industrial and Systems Engineering, National University of Singapore, 10 Kent Ridge Crescent, Singapore 119260

Abstract This paper describes the authors research on various heuristics in solving vehicle routing problem with time window constraints (VRPTW) to near optimal solutions. VRPTW is NP-hard problem and best solved to near optimum by heuristics. In the vehicle routing problem, a set of geographically dispersed customers with known demands and predened time windows are to be served by a eet of vehicles with limited capacity. The optimized routines for each vehicle are scheduled as to achieve the minimal total cost without violating the capacity and time window constraints. In this paper, we explore different hybridizations of articial intelligence based techniques including simulated annealing, tabu search and genetic algorithm for better performance in VRPTW. All the implemented hybrid heuristics are applied to solve the Solomons 56 VRPTW with 100-customer instances, and yield 23 solutions competitive to the best solutions published in literature according to the authors best knowledge. r 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Articial intelligence; Vehicle routing problems with time windows; Simulated annealing; Tabu search; Genetic algorithms

1. Introduction The vehicle routing problem with time window constraints (VRPTW) arises in a wide range of practical decision making problems, such as school bus routing, mail, newspaper delivery, fuel oil delivery and municipal waste collection. The VRPTW problem addressed in this paper belongs to the operational level in logistical decisions as classied by Hax and Candea (1984), which involves routing a eet of vehicles, with limited capacities, from a central depot to a set of geographically dispersed customers with known demands and predened time window constraints. The route cost of a vehicle is the total distance it travels, and the objective is to minimize the total cost of all routes with minimum number of vehicles without violating any constraints. Efcient routing and scheduling of vehicles can save millions of dollars for government and industries. Surveys on classications and applications of VRPTW can be found in Osman (1993a), Laporte (1992), Fisher (1993) and Bodin et al. (1983).
*Corresponding author. Tel.: +65-874-2127; fax: +65-779-1103. E-mail addresses: [email protected] (K.C. Tan), [email protected] (L.H. Lee), [email protected] (K. Ou).

It has been shown previously that nding a feasible solution for the traveling salesman problem (TSP), the simplest problem in vehicle routing, is an NP-complete problem. Therefore, solving the problem of VRPTW is more complicated since it involves serving many customers using multiple vehicles. Although optimal solutions to VRPTW can be obtained using exact methods, the computational time required to solve the VRPTW to optimality is prohibitive (Desrochers et al., 1992). Since heuristic methods often produce near optimal solutions in a reasonable amount of computational time, there are many researches currently working on the design of new heuristics for solving the VRPTW. A great amount of work has been done to the development of heuristics for VRPTW. For example, Christodes (1985), Fisher (1995), Federgruen and Simchi-Levi (1995), Bertsimas and Simichi-Levi (1996), Kontoravdis and Bard (1994), Kohl and Madsen (1997), Kohl et al. (1999), Thompson and Psaraftis (1993) or Kolen et al. (1987). These heuristics can be classied into four categories by Christodess denition: Constructive methods, route rst-cluster second method, cluster rstroute second methods and incomplete optimization methods. The saving algorithm (1964), heuristics of

0952-1976/02/$ - see front matter r 2002 Elsevier Science Ltd. All rights reserved. PII: S 0 9 5 2 - 1 9 7 6 ( 0 2 ) 0 0 0 1 1 - 8

826

K.C. Tan et al. / Engineering Applications of Articial Intelligence 14 (2001) 825837

Gaskel (1967), Yellow (1970) and Russel (1977) fall into the class of the constructive methods. The optimal partitioning heuristic (Beasley, 1983) and the sweep algorithm (Gillett and Miller, 1974) belong to the class of route rst-cluster second methods. There are three algorithms for the cluster rst-route second methods, which includes the two-phase method (Christodes et al., 1978), the generalized assignment heuristic (Fisher and Jaikumar, 1981) and the location-based heuristic (Bramel and Simchi-Levi, 1995). For the last category, it includes the method of cutting plane methods (Cornu! ejols and Harche, 1993) and minimum K-tree methods (Fisher, 1994). All the above algorithms are capable of giving feasible solutions for VRPTW, but the solutions produced may not be satisfactory because of their limited ability of searching the solution space. Heuristics search strategies based on genetic algorithm (GA), simulated annealing (SA) and tabu search (TS) (articial intelligence (AI) based algorithms) have been explored in recent years in order to improve the solutions in VRPTW, e.g., the works of Garcia et al. (1994), Potvin and Bengio (1996) on TS, Thangiah and Petrovic (1998), Thangiah et al. (1996) on GA and Thangiah (1999) on hybrid algorithms. All these AI-based algorithms can help to obtain better solutions for VRPTW by exploring much larger solution space. The work of Osman (1993b) in using SA and TS for VRPTW show that customer interchange methods guided by SA and TS produce solutions that are signicantly better than solutions obtained by competing heuristic methods. Heuristic search strategies based on GA (Thangiah and Nygard, 1992; Thangiah et al., 1993) were also proved to give good solutions for VRPTW. In this paper, we investigate the use of customer interchange method to improve solutions using local

search algorithm as well as hybrid SA and TS. Hybrid GA and local search method is also proposed in the paper to solve the representation problem in standard GA. For all these heuristics, the initial solution of VRPTW is obtained from push forward insertion heuristic (PFIH), which is a straightforward and constructive algorithm to obtain initial solutions (Solomon, 1987). The paper is arranged as follows: Section 2 gives a mathematical formulation of the VRPTW. Section 3 introduces the PFIH, l-interchange local search method, and local search with diversication. Section 4 describes the hybridization of TS and SA heuristics for VRPTW. Hybrid GA with local search technique is presented in Section 5. The overall computational results are listed and described in Section 6. Conclusions and future works are drawn in Section 7.

2. Mathematical formulation This section describes the notation and features that are common throughout the paper. Our formulation is based on the VRPTW model as shown in Fig. 1. The VRPTW consists of a set of identical vehicles, a central depot node, a set of customer nodes and a network connecting the depot and customers. It is assumed that there are N 1 customers, CAf0; 1; 2; y; Ng and K vehicles, V Af1; 2; y; Kg: Conveniently, the depot node is denoted as customer 0. Each arc in the network represents a connection between two nodes and also indicating the direction it travels. Each route starts from depot, visits customer nodes and then returns to depot. The number of routes in the network is equal to the number of vehicles used. One vehicle is dedicated to one route. A cost cij and a

4 5 1 3 6 2 Depot

8 9

7
Fig. 1. Model of VRPTW.

K.C. Tan et al. / Engineering Applications of Articial Intelligence 14 (2001) 825837

827

travel time tij are associated with each arc of the network. In Solomons 56 VRPTW 100-customer instances, all distances are represented by Euclidean distance, and the speed of all vehicles is assumed to be unity. That is, it takes one unit of time to travel one unit of distance. This assumption makes the problem simpler, because numerically the travel cost cij ; the travel time tij and the Euclidean distance between the customer nodes equal to each other. Each customer in the network can be visited only once by one of the vehicles. Every vehicle has the same capacity qk and each customer has a varying demand mi : qk must be greater or equal to the summation of all demands on the route traveled by vehicle k; which means that no vehicles can be overloaded. The time window constraint is denoted by a predened time interval, given an earliest arrival time and latest arrival time. The vehicles must arrive at the customers not later than the latest arrival time, while arrive earlier than the earliest arrival time, waiting occurs. Each customer also imposes a service time to the route, taking consideration of the loading/unloading time of goods. In Solomons instances, the service time is assumed to be unique regardless of the load quantity needed to be handled. Vehicles are also supposed to complete their individual routes within a total route time, which is essentially the time window of the depot. There are three types of decision variables in VRPTW. The decision variable xijk i; jAf0; 1; 2; y; Ng; kAf1; 2; y; Kg; iaj) is 1 if vehicle k travels from customer i to customer j; and 0 otherwise. The decision variable ti denotes the time when a vehicle arrives at the customer. And wi denotes the waiting time at node i: The objective is to design a network which satises all constraints, at the same time minimizing the total travel cost. The model is mathematically formulated below: Decision Variables: ti arrival time at node i; wi wait time at node i; xijk A0; 1; 0 if there is no arc from nodes i to j; and 1 otherwise. iaj: i; jA0; 1; 2; y; N: Parameters: K total number of vehicles, N total number of customers, Ci customer i; where iA1; 2; y; N; C0 central depot, yi any arbitrary real number, dij Euclidean distance between node i and node j; cij cost incurred on arc from node i to j; tij travel time between node i and j; mi demand at node i; qk capacity of vehicle k; ei earliest arrival time at node i;

latest arrival time at node i; li fi service time at node i; rk maximum route time allowed for vehicle k; minimize
N N K XXX i0 j0 k1

cij xijk

2:1

subject to:
K N XX k1 j1 N X j1 N X j1

xijk pK

for i 0;

2:2

xijk

xjik p1 kAf1; y; Kg; 2:3 2:4

for i 0
K N XX k1 j0 K N XX k1 i0 N X i0 N X j0

and

xijk 1 for iAf1; y; Ng;

xijk 1 for jAf1; y; Ng;

2:5

mi

xijk pqk

for kAf1; y; Kg;

2:6

N N XX i0 j0

xijk tij fi wi prk

for kAf1; y; Kg;

2:7

t0 w0 f0 0;
K N XX k1 i0

2:8

xijk ti tij fi wi tj for jAf1; y; Ng; 2:9 for iAf0; y; Ng: 2:10

ei pti wi pli

Formula (2.1) is the objective function of the problem. Constraint (2.2) species that there are maximum K routes going out of the depot. Eq. (2.3) makes sure every route starts and ends at the central depot. Eqs. (2.4) and (2.5) dene that every customer node can be visited only once by one vehicle. Eq. (2.6) is the capacity constraint. Eq. (2.7) is the maximum travel time constraint. Constraints (2.8)(2.10) dene the time windows. These formulas completely specify the feasible solutions for VRPTW.

3. Local search methods This section describes the route construction heuristic for obtaining an initial solution of VRPTW, the PFIH,

828

K.C. Tan et al. / Engineering Applications of Articial Intelligence 14 (2001) 825837

and the l-Interchange mechanism to generate neighboring solutions for improving the initial solution. 3.1. Push forward insertion heuristic (PFIH) The PFIHs for inserting customers into a route for the VRPTW was introduced by Solomon (1987). It is an efcient method for computing the cost of inserting a new customer into the current route. The procedure is easy and straightforward. The feasibility of inserting a customer into a route is checked by inserting the customer between all the edges in the current route and then select the edge that has the lowest additional insertion cost. The feasibility check examines all the constraints including time windows and load capacity. Only feasible insertion will be accepted. When the current route is full without accepting any new customer, PFIH will start a new route and repeat the procedure until all the customers are routed. PFIH serves the role of constructing route conguration for VRPTW, which is an efcient method to obtain feasible solutions. Detailed information of PFIH is available in Solomons paper (1987). 3.2. l-interchange local search method The l-Interchange local search method was introduced by Osman and Christodes (1994), which is used to improve the solution by interchanging customers between sets of routes. Based on some successfully solved problems (Osman, 1993a, b; Osman and Christodes, 1994; Thangiah et al., 1994), the l-Interchange local search method has been shown to be an effective neighborhood searching algorithm. The local search procedure is conducted by interchanging customer nodes between routes. For a chosen pair of routes, the searching order for the customers to be interchanged needs to be dened, either systematically or randomly. In this paper we only consider the cases l 1 and l 2; which means that maximum two customer nodes may be interchanged between routes. Based on the number of l; totally eight interchange operators can be dened, i.e., (0, 1), (1, 0), (1, 1), (0, 2), (2, 0), (2, 1), (1, 2), (2, 2). The operator (1, 2) on a route pair (Rp ; Rq ) indicates a shift of two customers from Rq to Rp and a shift of one customer from Rp to Rq : The other operators are dened similarly. For a given operator, the customers are considered sequentially along the routes. In both shift and interchange process, only improved solutions are accepted if the move results in the reduction of the total cost. There are two selection strategies proposed to select a candidate solution: (a) The rst-best (FB) strategy will select the rst solution S0 in Nl S; the neighborhood of the current solution that results in a decrease in cost;

(b) The global-best (GB) strategy will search all solution S 0 in Nl S; and select the one which will result in the maximum decrease in cost. From the description above, it can be understood intuitively that GB searches the solution neighborhood more thoroughly than FB, but needs longer time in process. 3.3. Local Search Method with Diversication (LS-DIV) Both of the above two selection strategies, e.g., GB and FB, are highly dependent upon the initial solution. They can only search very limited solution space in the vicinity of the initial solution and may get stuck in local optima having no means to get out of it. In order to explore larger solution space, a diversication mechanism is proposed in the local search method. When the local search method cannot produce any better solutions, it is believed that a local optimum has been reached. The current solution will then be diversied by randomly shifting and interchanging customers between randomly selected routes. As a result, the current solution jumps to a random neighbor. After several diversication processes, the current solution jumps to a new initial point in the solution space, and the l-interchange local search method is applied again. The only constraint of the diversication mechanism is that during the process only feasible moves will be accepted regardless of whether they lead to worse solutions or better solutions in terms of the total cost. The whole procedure is repeated, and there is no stopping criterion for this algorithm. It is generally believed that given enough time, this method can search big enough solution space and nd near optimal solutions. The steps of this algorithm are shown below: Heuristic 1: Local Search with Diversication (LS-DIV) LS-DIV-1: Obtain an initial solution from PFIH. LS-DIV-2: Use GB strategy to select the best solution candidate S 0 : LS-DIV-3: If {CS 0 oCS}then accept S0 and branch to LS-DIV-2. Else Branch to LS-DIV-4. LS-DIV-4: Diversify current solution M times. If{Termination criterion reached} Branch to LS-DIV-5. Else Branch to LS-DIV-2. LS-DIV-5: Terminate As mentioned, standard local search method is highly dependent on the initial solution and suffers the drawback of trapping in local optima. The LS-DIV

K.C. Tan et al. / Engineering Applications of Articial Intelligence 14 (2001) 825837

829

algorithm can be considered as an improved heuristic of the local search method, which helps to release such dependency and facilitates current solutions to escape from any local optima.

4. Hybrid algorithm of simulated annealing and tabu search SA and TS are two well-known Meta heuristics in solving optimization problems. In this paper, we propose a hybrid heuristic algorithm by combining SA and TS to solve the problem of VRPTW. 4.1. Simulated annealing Kirkatrick et al. (1983) started the work of solving combinatorial optimization problem using SA, which uses a stochastic approach to guide the search. Besides accepting better solutions, SA also allows the search to proceed to a neighboring state even if the move causes the value of the objective function to become worse. SA directs the search in an optimization problem in the following way: If a move to a neighborhood S0 decreases the objective function value, or leaves it unchanged, then the move is always accepted. More precisely, the solution S 0 is accepted as the new solution if Dp0; where D CS 0 CS; and CS is the cost of the solution. In order to allow the search to escape from a local optimum, moves which increase the objective function value may also be accepted according to a probability function eD=T > y if D > 0; 4:1

1990, 1997). TS stores the most recently made moves or visited solutions in a tabu list. Attempts that reverse the moves or reproduce the solutions in the tabu list will be denied and marked as Tabu. However, this restriction can be realized by an aspiration criterion only if the attempt of move will lead to a new global best solution. The lifetime of a tabu staying in the tabu list is controlled by the tabu list size, and the rst-inrstout rule is often used in refreshing the tabu list. There are two kinds of structures for tabu list. The rst one records the recent moves of individual nodes and set them as Tabu. The structure is described below fR1 ; Node 1; R2 ; Node 2; R3 ; Node 3; yg: 4:2

Each pair of the numbers such as {R1, Node 1} indicates that Node 1 is moved out from route R1 ; and likewise for the other pairs. Any move that attempts to shift Node 1 back to its original route R1 is prohibited as long as {R1, Node 1} is still in the tabu list. The second structure records the whole route information when the shift of nodes occurs instead of only recording the node number being shifted. If any of the nodes in this route is shifted out in the recent move, the whole route is recorded as a Tabu. The elements of this structure are strings of node numbers representing recently visited routes, e.g., f5-9-8-7-2; 12-4-6-3-21; 11-3-13-32-15; yg : 4:3

where T is the parameter of temperature, and y is a random number between [0, 1]. The value of T varies gradually from a relatively high temperature to a small value close to zero, which is often controlled by a linear cooling schedule for which the temperature is decreased linearly. The temperature T will be reset to a high value after the occurrence of a special neighborhood without accepting any moves. SA has no ending criterion and has been statistically proved to be capable of nding the optimal solution (Davis, 1991). However, SA may get caught in repetition of moves, e.g., a succession of moves that result in cycles. This problem of SA can be disastrous because of the large consumption of computational time. In order to avoid moves that result in cycles and force the search to explore other regions, TS is introduced in the following section. 4.2. Tabu search TS was introduced as a memory structure to support and encourage its non-monotonic search (Glover, 1989,

The tabu restriction imposed by this structure is similar to the rst structure. Any move is prohibited if it attempts to produce the same route that has been encountered before. Obviously, the second structure provides a more accurate memory for recording the recently made moves, although it takes longer computational time for comparison. In this paper, the second structure is adopted as the tabu list and its length is empirically set to 10. 4.3. Hybrid heuristic of SA and TS In the hybrid heuristic of SA and TS, a nonmonotonic cooling schedule is used to decrease the temperature after each accepted move, which is given as p Tk1 Tk =1 t Tk : 4:4 The starting temperature Ts is set to 50, and the time constant t is set to 0.3. When the temperature is high, the probability of accepting worse move is also high. When the temperature is decreased gradually, the chance of accepting worse move is also reducing and resulting the search approaches the thermal equilibrium point. When no more moves can be accepted, the local optimum is assumed to be reached and the whole cooling process is repeated by resetting

830

K.C. Tan et al. / Engineering Applications of Articial Intelligence 14 (2001) 825837

the temperature to a high initial value as given by Tr Max Tr =2; Tb ; 4:5 where Tr is the reset temperature and Tb is the temperature at which the best current solution is found. The candidate moves are searched systematically in the hybrid heuristic, which means that the search of move is started from the rst position of the rst route onwards. During the process, if the move is not a tabu and satises the SA criterion, it will be accepted and then the search will be restarted from the beginning of a new current solution after updating the tabu list and SA parameters. It follows the idea of FB local search method, but with a tabu structure added to remove any cycling and an SA criterion attached to guide the stochastic search. The hybrid algorithm of SATabu is described below Ts Starting temperature of SA=50, Tr Reset temperature of SA, originally equal to Ts ; Tb Temperature at which the current best solution is found, Tk Temperature of the current solution, S Current solution, Sb Current best solution, R Number of resets done, t Time constant in the range [0, 1].

SATabu-8: If {R is greater than a preset number} Terminate the program. Else Branch to SATabu-4.

5. Genetic algorithm 5.1. Standard GA GA was introduced by Holland (1975), which states that even in large and complicated search spaces, given certain conditions on the problem domain, GA would tend to converge to a solution that is globally optimal or nearly so. GA has a standard structure and procedure, including representation of chromosomes, creation of initial population, selection, reproduction and local search improvement (LSI). 5.1.1. Representation Most of the GA applications use binary strings to represent chromosomes. Unlike these applications, we use an integer string of length N as the representation of chromosome, where N is the number of customers in question. Each gene in the chromosome is the customer node number assigned to them, and the sequence of the genes is the order of visiting these customers. For example, a chromosome representation can be shown as 2 5 8 10 12 9 3 7 1 4 11: 5:1 This chromosome represents one feasible solution of 12 node VRPTW. In this representation, there is no depot node number 0, the delimiters, since such delimiters in a chromosome greatly restrain the validity of children chromosomes reproduced by crossover operation. To decode this chromosome into route conguration, we simply insert the genes into new routes sequentially in a similar way to PFIH. When a route is full, a new route starts until the last gene is inserted. 5.1.2. Creation of initial population In VRPTW, a portion of the population members can be initialized using the PFIH initial solution and its random neighborsANS; while the remaining chromosomes can be produced randomly. The population size is kept at 1000 in our study, and the total generation number is set at 5001000, a compromise between computational time and the results obtained. 5.1.3. Genetic operators In our problem, the tness value associated with each chromosome is their respective total cost in the route conguration. Chromosomes with good tness value (lower total cost) will have higher chance to be selected as parents for mating and reproduction. Crossover is the

Heuristic 2: Hybrid simulated annealing and Tabu search (SaTabu) SATabu-1: Obtain an initial solution S using PFIH. SATabu-2: Initialize Tabulist to zero and all the SA parameters. SATabu-3: Set the iteration counter m 0 and the reset counter R 0: SATabu-4: Generate systematically a S 0 in Nl S: And compute D CS 0 CS: SATabu-5: If {S0 is a Tabu} Branch to SATabu-4 and select the next move. SATabu-6: Accept or reject S0 according to the SA criterion. If {S0 is accepted} S S0 ; update Tabulist and other SA parameters. If {CSoCSb } Update Sb S: Branch to SATabu-7. Else Branch to SATabu-4 and select the next move. SATabu-7: If {the whole neighborhood is searched without any accepted move} Reset temperature, R R 1: Branch to SATabu-4.

K.C. Tan et al. / Engineering Applications of Articial Intelligence 14 (2001) 825837

831

key operator to generate new individuals in the population. The mating pool has N=2 pairs of chromosomes and the crossover operation is applied to each pair with the crossover probability pc ; the value of pc is in general >0.6. The PMX crossover operator is used in this paper (Thangiah, 1995). Mutation serves the role of bringing random, unrelated traits into the present population and increasing the variance of the population. The mutation process involves randomly swapping nodes in the same chromosome. There is a probability associated with mutation, pm in the range of [0,1]. Earlier research has shown that excessive pm drives the GA into convergence sooner than necessary, resulting in undesirable local optimal solutions. On the contrary, small pm produces the opposite results, e.g., intolerably slow convergence. An adaptive mutation probability scheme is used in this paper, which adapts pm to the standard deviation of the population s PN1 2 % i0 xi x S ; 5:2 N 1 where N is the population size, xi is the tness value of individual i and x is the average tness of the % population. If S is greater or equal to a threshold value MINPOPDEV=0.5, a minimum pm pmin 0:06 is used, otherwise pm pmin MINPOPDEVS: 5.1.4. Local search improvement (LSI) After a new generation is produced via the above reproduction process, a portion of chromosomes is decoded into their route conguration, and undergoes a l-interchange procedure using local search method to improve the solutions. The improved solutions are encoded back to the chromosome format. This process is called LSI, which is a supplementary solution improvement process. Note that not all of the chromosomes will undergo the LSI process due to the consumption of time. In the algorithm, there is a LSI probability parameter ph ; which is set to 0.5. When chromosomes are undergoing the LSI process, the FB linterchange procedure is used since it is less timeconsuming than the method of GB. 5.2. Hybrid GA As described above, the chromosome representation adopted in this paper is in the format of an integer string, e.g., it could be in the form of 2 5 8 10 12 9 3 7 6 1 4 11 : 5:3

be decoded into route conguration as 2 5 8 10 12 9 3 7 6 1 4 11 ; 5:4

where each portion in the parentheses represents one route. Clearly, although a feasible solution can be guaranteed, the solution may not be the best one that the chromosome string can represent. For the same chromosome string in (5.3), it may also represent many feasible solutions as long as all the constraints are not violated, such as
2 2 2 5 8 10 5 8 10 5 8 10 12 12 9 9 3 7 6 1 4 3 7 6 1 4 6 11 11 5:5

12 9 3 7

1 4 11:

The only difference between these representations is the grouping of customer nodes. With different groupings, one chromosome can represent different solutions. To search better grouping (better tness value) for each chromosome, a new hybrid GA is proposed in this paper which associates the grouping information with each chromosome, e.g., the chromosome representation may be given as
2 5 8 10 2 5 8 10 2 5 8 10 12 9 3 7 12 9 3 7 12 9 3 7 6 1 4 11 4 4 6 1 4 11 3 6 4 3 5:6

6 1 4 11 5 5 2:

This chromosome can be decoded into route conguration by sequentially pushing the customer nodes into new routes, which is a very greedy constructive method. By using the method, the above chromosome string may

In each of these new representations, the second part inside the parentheses is the grouping information. For example, 3 6 3 indicates that the rst 3 customer nodes are in the rst route, the next 6 customer nodes are in the second route, and so on. The tness value can then be calculated according to the grouping information easily. The appropriate selection of grouping for chromosomes is the major focus in this hybrid GA. Every chromosome has an initial grouping information using the greedy constructive method of sequential pushing as used in the standard GA. A local search method is then used to search new groupings for each chromosome, which is conducted by changing the grouping information until one feasible better grouping is accepted. This searching procedure is similar to the l-interchange local search method as described in Section 3, and the only changes made here is on the grouping information. For example, if the original grouping is 6 5 1 ; we can check the new grouping combinations such as 5 6 1 ; 5 5 2 ; 4 5 3 and so on. The searching order and operators can be dened easily. Each of the new groupings will be checked for validity rst, and only feasible groupings will be under consideration. The search will stop when a new better grouping is found as similar to the strategy of FB selection. If there is no new grouping found when the searching time is out, the grouping information remains unchanged.

832

K.C. Tan et al. / Engineering Applications of Articial Intelligence 14 (2001) 825837

After the grouping update process, chromosomes will go through the LSI for further ne-tuning. The whole improvement procedure including grouping update and LSI is repeated several times, and then the population will undergo the same selection, reproduction process as in the standard GA. The hybrid GA procedure is shown below, where P1 and P2 are two preset numbers for the groupings. Heuristic 3: Hybrid genetic algorithm (HGA) HybridGA-1: Initialize chromosomes and set the iteration counter N 0: HybridGA-2: Get initial grouping information for each chromosome. HybridGA-3: Set the regrouping counter M: HybridGA-4: Update chromosome-grouping information using FB local search. HybridGA-5: Undergo LSI Process, M M 1: HybridGA-6: If (M > P1) Branch to HybridGA-7. Else Branch to HybridGA-4. HybridGA-7: GA operations: selection, crossover and mutation. N N 1: HybridGA-8: If (N > P2) Branch to HybridGA-9. Else Branch to HybridGA-2. HybridGA-9: Terminate. The advantages of using the proposed hybrid GA can be summarized below:
*

with diversication (LS-DIV); Hybrid SA with Tabu (SATabu) and Hybrid Genetic Algorithm (HGA). The heuristics were applied to six data VRPTW sets including R1, C1, RC1, R2, C2 and RC2 as generated by Solomon (Solomon, 1987) consisting of 100 customers with Euclidean distance. In these problems, the travel time between the customers is equal to the corresponding Euclidean distance. The data consist of geographical and temporal differences in addition to differences in demands for the customers. Each of the problems in these data sets has 100 customers. The eet size to service them varies between 2 and 21 vehicles. The VRPTW problems generated by Solomon incorporate many distinguishing features of vehicle routing with two-sided time windows. The problems vary in eet size, vehicle capacity, travel time of vehicles, spatial and temporal distribution of customers, time window density (the number of demands with time windows), time window width, percentage of time constrained customers and customer service times. Problem sets R1, C1 and RC1 have narrow scheduling horizon. Hence, only a few customers can be served by the same vehicle. Conversely, problem sets R2, C2 and RC2 have large scheduling horizon, and more customers can be served by the same vehicle. Problems belonging to C categories are clustered data, meaning nodes are clustered either geographically or in term of time windows. Problems from R categories are uniformly distributed data and those from RC categories are hybrid problems having the features of both C and R categories. 6.1. Comparison results of competing heuristics Performance of the three proposed heuristic methods in this paper have been compared with another nine competing methods as shown in Table 1. The explanation of the table is shown below: (a) Data sets, from C1 to RC2, are compared instead of comparing every instance; (b) In each grid there are two numbers representing the average vehicle number and average travel distance, respectively (the upper one for vehicle number and lower one for distance). For example, in class C1, the number pair (10.00, 828.45) means that over the 9 instances in C1, the average vehicle number deployed is 10, and the average travel distance is 828.45; (c) The last row is the total accumulated sum, indicating the total number of vehicles used in all 56 instances and the total travel distance. The results obtained from the comparison can be summarized below (a) In category R1, LS-DIV achieves the best average results. SATabu produces best solutions in R2 and

The chromosome representation becomes much more exible. One-to-Many replaces One-to-One representation; By choosing from the One-to-Many representation, each chromosome is best represented in terms of the cost value; Fine-tuning procedure including grouping update and LSI plays an important role in this hybrid GA. Since PMX has been shown to be a disruptive crossover method for VRPTW problems in which the order is important, the standard GA is difcult to produce optimal solutions since the good relations in chromosomes will hardly be maintained after normal crossover operation; The ne-tuning procedure is in charge of rearranging the customer order in chromosomes and the GA operations serve the role of interchanging genetic materials between chromosomes as usual.

6. Computational results In this paper, three heuristics are proposed to solve the VRPTW, e.g., l-interchange local search method

K.C. Tan et al. / Engineering Applications of Articial Intelligence 14 (2001) 825837 13.3 1366.62 10.0 851.05 3.2 620.12 4.4 985.69 5.2 1108.5 HGA 470 57903

833

Bachem et al. (1997)

12.25 1264.24

2.91 1100.33

11.75 1414.63

3.38 1341.35

10.00 828.38

3.00 591.88

414 61523

HGA in RC1 and RC2. It is also noticed that in RC2, both results from SATabu and HGA are better than those from competing heuristics. In all the R class data sets where the Solomon instances have uniformly distributed customers, our heuristics give satisfactory results. For SATabu, it favors those VRPTW instances having uniformly distributed customers and large scheduling horizon in R2 and RC2. Likewise, HGA is shown to generate good solutions for data instances in RC class. (b) One observation from the results is the average vehicle eet size. Our solutions usually employ more vehicles. However, as stated in the beginning of this paper, the only aim of our algorithm is to reduce the total distance traveled by the vehicles. The size of the eet does not impose any additional cost, so that it is possible for our algorithm to reduce the total distance by increasing the number of vehicles. (c) The cumulated total travel distance given by the three heuristics are fairly good as compared with other heuristics. Among the total 12 heuristics, SATabu is ranked 2nd, HGA is ranked 4th and LSDIV is ranked 7th. The solutions for all the instances are listed in Appendix A. The results presented here are the best solutions collected after running the program independently for many times for each instance. Although we only marked down the best result for each benchmark problem, the standard deviation of the results is low, around 13%, which shows that the solutions produced from our proposed heuristics are consistent. 6.2. Comparison with the best-known results A comparison between our heuristics and the bestknown results in literature according to the authors best knowledge is shown in Appendix A. Those results are collectively gathered from different papers in literature covering the most recent research on the optimization of VRPTW in using Solomons data sets. All the references are clearly presented in Appendix A, and those competitive results are highlighted. Most of the better/ equivalent solutions are observed to come from the categories of C1, R2 and RC2. Table 2 shows the average travel distance for each class of data similar to Table 1, the percentage increase of our heuristics against the best-known results according to our best knowledge as well as the number of solutions obtained from our proposed heuristics that are better than or equivalent to the best solutions (indicated by a fraction, e.g., 3/9 means 3 good solutions out of 9 instances). As can be seen, the average performance over the six data sets against the reference results for the three different heuristics are shown in the table. In each of the

SATabu

13.1 1213.16

13.2 1220

12.7 1415.62

5.6 1120.37 4.8 1157.1 3.25 1144.43 3.25 1229.54 3.38 1165.62 3.40 1360.60 3.40 1476.10 3.38 1401.50 3.38 1398.00 3.62 1139.79 RC2

10.0 841.92

3.3 612.75

LS-DIV

13.3 1197.27

10.0 850.71

4.5 979.53

4.6 952.3

13 1413.1

Homberger and Gehring (1999)

3.1 646.2

11.92 1228.06

11.63 1392.57

10.00 828.38

3.00 589.86

2.73 969.95

Chiang and Russell (1997)

2.17 1204.19

11.88 1397.44

10.00 828.38

3.00 591.42

2.73 986.32

Taillard et al. (1996)

12.25 1216.70

11088 1367.51

10.00 828.45

3.00 590.30

3.00 995.38

Potvin and Bengio (1996)

12.60 1296.83

3.00 1117.70

12.10 1446.20

10.00 838.00

3.00 589.90

Potvin et al. (1996)

12.60 1294.70

3.10 1185.90

12.60 1465.00

10.00 861.00

3.00 602.50

Chiang and Russell (1996)

12.50 1308.82

2.91 1166.42

12.38 1473.90

10.00 909.80

3.00 684.10

Table 1 Comparison among different heuristics

12.66 1317.00

2.91 1167.00

Russell (1995)

12.38 1523.00

10.00 930.00

3.00 681.00

Rochat and Taillard (1995)

12.58 1197.42

12.38 1369.48

10.00 828.45

3.00 590.32

3.09 954.36

Prob class

RC1

All

R1

R2

C1

C2

427 57120

424 65827

422 65201

427 64679

422 62572

416 57993

411 58502

406 57876

466 58529

470 57799

834

K.C. Tan et al. / Engineering Applications of Articial Intelligence 14 (2001) 825837

Table 2 Percentage increase and number of good solutions Prob class C1 C2 R1 R2 RC1 RC2 All Best result 827.5 589.9 1174.7 942.0 1354.1 1081.3 56108 LS-DIV 850.71 646.2 1197.27 979.53 1413.1 1157.1 58529 +2.8% +9.5% +1.9% +4.0% +4.3% +7.0% +4.3% 3/9 1/8 2/12 4/11 2/8 1/8 13/56 SATabu 841.92 612.75 1213.16 952.3 1415.62 1120.37 57799 +1.7% +3.9% +3.3% +1.1% +4.5% +3.6% +3.0% 5/9 2/8 2/12 4/11 0/8 3/8 16/56 HGA 851.05 620.12 1220 985.69 1366.62 1108.5 57903 +2.8% +5.1% +3.8% +4.6% +0.9% +2.5% +3.2% 3/9 1/8 2/12 4/11 2/8 3/8 15/56

Note: Percentage=(mGA resultbest result)/best result.

Table 3 Comparison between optimal solutions and our heuristic solutions Data set C101 C102 C106 C107 C108 R101 R102 Optimum 10/827.3 10/827.3 10/827.3 10/827.3 10/827.3 18/1607.7 17/1434.0 LS-DIV 10/828.937 10/831.734 10/840.482 10/828.937 10/853.339 19/1648.86 18/1482.77 +0.2% +0.5% +1.6% +0.2% +3.1% +2.5% +3.4% SATabu 10/828.937 10/828.937 10/828.937 10/828.937 10/830.939 19/1655.03 19/1517.06 +0.2% +0.2% +0.2% +0.2% +0.4% +2.9% +5.8% HGA 10/828.937 10/832.672 10/836.653 10/828.937 10/861.999 19/1660.33 18/1486.82 +0.2% +0.6% +1.1% +0.2% +4.2% +3.3% +3.7%

data sets, our results are closely comparable and the overall distance gap between our heuristics and the best results is ranging from 3% to 4.3% which is acceptable. 6.3. Comparison with optimal solutions In this section, the optimal solutions for some of the VRPTW instances reported in (Desrochers et al., 1992) are compared with our heuristic results in Table 3. The percentage increase of the travel distance is also included. For these seven optimal solutions, our heuristics have found close results especially for the C1 class instances. 6.4. Computational time

Table 4 Average time needed for each method Heuristics Time (s) Standard deviation LS-DIV 158 2.3% SATabu 275 1.2% HGA 1340 1.8%

computational time for each run. As can be seen, HGA takes the longest time to run as expected because of its large population size and its gradual evolutionary manner. LS-DIV and SATabu take shorter time to run and offer reasonably good results.

7. Conclusions and future works Our three heuristics use different approach to solve VRPTW, and their computational time is different consequently. Table 4 shows the average time and standard deviation for the heuristics. For LS-DIV, each instance is diversied 10 times after the current solution is stuck in local optima. For SATabu, the total temperature reset is 10 times. HGA has a population size of 1000 and a generation size of 500. All the algorithms are written in C++ programming language, and were run in a Pentium II 330 Mhz PC. There were at least 10 independent runs for each of the instances in each algorithm. The time needed to nd the solutions is consistent and with only minor difference in the In this paper, the usage and implementation of l-interchange local search method and its modied heuristic LS-DIV have been studied. The heuristic is simple to implement and the obtained current solution is capable of avoiding local optima via the implementation of diversication. A hybrid heuristics of SA and TS has also been proposed, which is capable of exploring larger search space by overcoming the problem of local optima or cycling of moves. In addition, a new hybrid GA has been proposed in this paper, which adopts a new chromosome structure containing more information about the vehicle route conguration. By combining

K.C. Tan et al. / Engineering Applications of Articial Intelligence 14 (2001) 825837

835

the new structure with local search approaches, each chromosome representation can be ne-tuned for better routing solutions. Results obtained are encouraging and show the effectiveness of the proposed hybrid heuristics. An extensive computational study on a set of benchmark problems based on the Solomons 56 VRPTW with 100-customer instances has been performed. All the results obtained from our heuristics have been compared with the best-known results in literature according to our best knowledge. It has been shown that our heuristics produce competitive solutions as well as achieve good averaged results.

To reduce the needed computational time of running the proposed heuristics, a parallel and distributed GA is currently being developed using the Internet and Peerto-Peer computing technology. Such parallel computing approach should also increase the efciency and accuracy of our heuristics by employing migration strategy in the algorithm. Progress and results will be reported in due course.

Appendix A See Table 5.

Table 5 Detail results of the heuristics Prob class Reference Best-known NV C101 C102 C103 C104 C105 C106 C107 C108 C109 C201 C202 C203 C204 C205 C206 C207 C208 R101 R102 R103 R104 R105 R106 R107 R108 R109 R110 R111 R112 R201 R202 R203 R204 R205 R206 R207 R208 R209 R210 R211 Desrochers et al., 1992 Desrochers et al., 1992 Rochat and Taillard, 1995 Rochat and Taillard, 1995 Potvin and Bengio, 1996 Desrochers et al., 1992 Desrochers et al., 1992 Desrochers et al., 1992 Potvin and Bengio, 1996 Potvin and Bengio, 1996 Potvin and Bengio, 1996 Rochat and Taillard, 1995 Potvin and Bengio, 1996 Potvin and Bengio, 1996 Potvin and Bengio, 1996 Rochat and Taillard, 1995 Rochat and Taillard, 1995 Desrochers et al., 1992 Desrochers et al., 1992 Thangiah et al., 1994 Rochat and Taillard, 1995 Rochat and Taillard, 1995 Rochat and Taillard, 1995 Homberger and Gehring, 1999 Taillard et al., 1996 Chiang and Russell, 1997 Rochat and Taillard, 1995 Homberger and Gehring, 1999 Rochat and Taillard, 1995 Homberger and Gehring, 1999 Homberger and Gehring, 1999 Homberger and Gehring, 1999 Homberger and Gehring, 1999 Homberger and Gehring, 1999 Thangiah et al., 1994 Rochat and Taillard, 1995 Homberger and Gehring, 1999 Thangiah et al., 1994 Homberger and Gehring, 1999 Homberger and Gehring, 1999 10 10 10 10 10 10 10 10 10 3 3 3 3 3 3 3 3 18 17 13 10 14 12 10 9 12 11 10 10 4 3 3 2 3 3 3 2 3 3 2 TD 827.3 827.3 828.26 824.78 828.94 827.3 827.3 827.3 828.94 591.56 591.56 591.17 590.60 588.88 588.49 588.29 588.32 1607.7 1434.0 1207 982.01 1377.11 1252.03 1120.85 968.59 1013.16 1080.36 1099.46 953.63 1252.37 1198.45 942.64 854.88 1013.47 833 814.78 731.23 855 955.39 910.09 LS-DIV NV 10 10 10 10 10 10 10 10 10 3 3 3 3 4 3 3 3 19 18 15 12 15 12 12 11 13 11 12 10 6 6 5 5 5 4 3 4 3 5 4 TD 828.937 831.734 869.178 900.467 828.737 840.482 828.937 853.339 875.312 591.557 612.249 620.772 655.921 680.954 674.179 612.251 720.953 1648.86 1482.77 1221.06 1032.03 1378.11 1288.22 1123.7 989.57 1027.49 1108.07 1086.36 981.09 1282.41 1147.53 971.985 813.712 1015.99 1010.72 921.552 778.381 945.276 1003.26 884.703 SATabu NV 10 10 10 10 10 10 10 10 10 3 3 3 4 3 4 3 4 19 19 13 10 16 13 12 10 13 11 12 10 8 6 5 4 5 5 4 3 3 5 3 TD 828.937 828.937 860.39 894.822 828.937 828.937 828.937 830.939 849.034 591.557 591.557 628.124 601.23 598.764 653.788 618.023 618.673 1655.03 1517.06 1234.43 1024.38 1382.71 1271.11 1116.04 994.876 1129.79 1119.53 1091.11 1020.52 1198.15 1077.66 933.286 826.19 1049.04 974.5 875.757 769.256 912.47 961.179 896.108 HGA NV 10 10 10 10 10 10 10 10 10 3 3 3 4 4 3 3 3 19 18 13 11 16 13 12 10 14 11 12 10 6 6 5 4 5 5 3 3 3 5 4 TD 828.937 832.672 859.779 893.225 828.937 836.653 828.937 861.999 890.92 591.557 613.478 646.776 605.345 631.21 648.915 609.157 614.951 1660.33 1486.82 1272.14 1073.63 1380.44 1283.06 1117.81 995.361 1101.37 1162.19 1093.76 1012.85 1243.18 1188.91 1050.03 800.361 1056.54 984.648 920.269 770.691 902.67 1037.58 887.476

836 Table 5 (continued) Prob class Reference

K.C. Tan et al. / Engineering Applications of Articial Intelligence 14 (2001) 825837

Best-known NV TD 1642.82 1540.97 1110 1135.83 1637.15 1395.37 1230.54 1139.82 1249 1164.25 1060.45 799.12 1302.42 1158.81 1082.32 833.97

LS-DIV NV 14 15 11 12 15 14 12 11 4 5 5 4 6 5 5 5 TD 1677.68 1512.02 1210.29 1252.41 1626.47 1465.07 1338.69 1221.48 1308.31 1251.52 1142.39 848.05 1463.7 1148.33 1171.52 922.566

SATabu NV 14 14 12 10 15 13 12 12 5 8 7 4 6 5 5 5 TD 1677.93 1569.73 1196.12 1264.73 1648.88 1480.06 1303.36 1184.68 1303.64 1151.46 1032.7 865.513 1426.09 1162.6 1078.65 942.78

HGA NV 15 15 13 10 16 15 13 10 5 5 5 4 8 5 5 5 TD 1658.96 1514.85 1149.86 1173.47 1585.34 1403.1 1290.76 1157.20 1354.96 1257.48 1063.77 899.347 1236.18 1133.86 1068.73 854.749

RC101 RC102 RC103 RC104 RC105 RC106 RC107 RC108 RC201 RC202 RC203 RC204 RC205 RC206 RC207 RC208

Chiang and Russell, 1997 Chiang and Russell, 1997 Thangiah et al., 1994 Rochat and Taillard, 1995 Homberger and Gehring, 1999 Chiang and Russell, 1997 Taillard et al., 1997 Taillard et al., 1997 Thangiah et al., 1994 Taillard et al., 1997 Homberger and Gehring, 1999 Homberger and Gehring, 1999 Homberger and Gehring, 1999 Thangiah et al., 1996 Thangiah et al., 1996 Rochat and Taillard, 1995

15 13 11 10 13 12 11 10 4 4 3 3 4 3 3 3

References
Bachem, A., Bodmann, M., Bolz, G., Emden-Weinert, T., Erdmann, + A., Kiahaschemi, M., Monien, B., Promel, H.J., Schepers, J., + Schrader, R., Schulze, J., Tschoke, S., 1997. Verbundprojekt PARALOR: Parallele Algorithmen zur Wegeoptimierung in Flugpaanung und Logistik. Report No. 97-269, Beitrag zur Statustagung des BMBF HPSC 97, Angewandte Mathematik und + + Informatik, Universit. t Koln, Koln. a Beasley, J., 1983. Route rst-cluster second methods for vehicle routing. Omega 11, 403408. Bertsimas, D., Simchi-Levi, D., 1996. The new generation of vehicle routing research: robust algorithms addressing uncertainty. Operations Research 44, 286304. Bodin, L., Golden, B., Assad, A., Ball, M., 1983. The state of the art in the routing and scheduling of vehicles and crews. Computers and operations Research 10 (2), 63211. Bramel, J., Simchi-Levi, D., 1995. A location based heuristic for general routing problems. Operations Research 43, 649660. Chiang, W.-C., Russell, R.A., 1996. Simulated annealing metaheuristics for the vehicle routing problem with time windows. Annals of Operations Research 63, 327. Chiang, W.-C., Russell, R.A., 1997. A reactive tabu search metaheuristics for the vehicle routing problem with time windows. INFORMS Journal on computing 9, 417430. Christodes, N., 1985. Vehicle routing. In: Lawler, E.L., Lenstra, J.K., Rinnooy Kan, A.H.G., Shmoys, D.B. (Eds.), The Traveling Salesman Problem: a Guided Tour of Combinatorial Optimization. Wiley, New York, pp. 431448. Christodes, N., Mingozi, A., Toth, P., 1978. The vehicle routing problem. In: Christodes, N., Mingozzi, A., Toth, P., Sandi, C. (Eds.), Combinatorial Optimization. Wiley, New York, pp. 318338. Cornu! jols, G., Harche, F., 1993. Polyhedral study of the capacitated e vehicle routing problem. Mathematical Programming 60, 2152. Davis, L., 1991. Handbook of Genetic Algorithms. Van Nostrand Reinhold, New York.

Desrochers, M., Desrosiers, J., Solomon, M., 1992. A new optimization algorithm for the vehicle routing problem with time windows. Operations Research 40 (2), 342354. Federgruen, A., Simchi-Levi, D., 1995. Analytical analysis of vehicle routing and inventory routing problems. In: Ball, M.O., Magnanti, T.L., Monma, C.L., Nemhauser, G.L. (Eds.), Handbooks in Operations Research and Management Science, the Volume on Network Routing. North-Holland, Amsterdam, pp. 297373. Fisher, M., 1993. Vehicle Routing. Handbooks on Operations Research and Management Science, Volume on Networks, North-Holland, Amsterdam, in preparation. Fisher, M.L., 1994. Optimal solution of vehicle routing problems using minimum k-tree. Operations Research 42, 626642. Fisher, M.L., 1995. Vehicle routing. In: Ball, M.O., Magnanti, T.L., Monma, C.L., Nemhauser, G.L. (Eds.), Handbooks in Operations Research and Management Science, the Volume on Network Routing. North-Holland, Amsterdam, pp. 133. Fisher, M.L., Jaikumar, R., 1981. A generalized assignment heuristic for vehicle routing. Networks 11, 109124. Garcia, B.-L., Potvin, J.-Y., Rousseau, J.-M., 1994. A parallel implementation of the tabu search heuristic for vehicle routing problems with time window constraints. Computers & Operations Research 21 (9), 10251033. Gaskel, T.J., 1967. Bases for vehicle eet scheduling. Operational Research Quarterly 18, 281295. Gillett, B.E., Miller, L.R., 1974. A heuristic algorithm for the vehicle routing dispatch problem. Operational Research 22, 340349. Glover, F., 1989. Tabu Search-Part I. ORSA Journal on Computing 1 (3), 190206. Glover, F., 1990. Tabu search-part II. ORSA Journal on Computing 2 (1), 432. Glover, F., 1997. Manuel Lagun. In: Tabu Search. Kluwer Academic Publishers, Dordrecht. Hax, A.C., Candea, D., 1984. Production and Inventory Management. Prentice-Hall, Englewood Cliffs, NJ. Holland, J.H., 1975. Adaptation in Natural and Articial Systems. University of Michigan Press, Ann Arbor.

K.C. Tan et al. / Engineering Applications of Articial Intelligence 14 (2001) 825837 Homberger, J., Gehring, H., 1999. Two evolutionary metaheuristics for the vehicle routing problem with time windows. INFOR 37, 297318. Kirkatrick, S., Gelatt, C.D., Vecchi, P.M., 1983. Optimization by simulated annealing. Science 220, 671680. Kohl, N., Madsen, O.B.G., 1997. An optimization algorithm for the vehicle routing problem with time windows based on Lagrangean relaxation. Operations Research 45 (3), 395406. Kohl, N., Desrosiers, J., Madsen, O.B.G., Solomon, M.M., Soumis, F., 1999. 2-path cuts for the vehicle routing problem with time windows. Transportation Science 33 (1), 101331116. Kolen, A.W.J., Rinnooy Kaan, A.H.G., Trienekens, H.W.J.M., 1987. Vehicle routing with time windows. Operations Research 35 (2), 266331273. Kontoravdis, G., Bard, J.F., 1994. Improved Heuristics for the Vehicle Routing Problem with Time Windows, Technical Report, Department of Mechanical Engineering, University of Texas, Austin, TX. Laporte, G., 1992. The vehicle routing problem: an overview of exact and approximate algorithms. Europe Journal of Operational Research 59, 345358. Osman, I.H., 1993a. Vehicle routing and scheduling: application, algorithms and developments. Proceedings of the International Conference on Industrial Logistics, Rennes, France. Osman, I.H., 1993b. Metastrategy simulated annealing and tabu search algorithms for the vehicle routing problems. Annals of Operations Research 42, 421451. Osman, I.H., Christodes, N., 1994. Capacitated clustering problem by hybrid simulated annealing and tabu search. International Transaction in Operational Research 1 (3) 317336. Potvin, J.-Y., Bengio, S., 1996. The vehicle routing problem with time windowspart II: genetic search. INFORMS Journal on Computing 8, 165172. Potvin, J.Y., Kervahut, T., Garcia, B.L., Rousseau, J.M., 1996. The vehicle routing problem with time windows; part I: tabu search. INFORMS Journal on Computing 8, 158164. Rochat, Y., Taillard, E.D., 1995. Probabilistic diversication and intensication in local search for vehicle routing. Technical Report CRT-95-13, Centre de recherche sur les transports, Universit! de e Montr! al, Montr! al. e e Russel, R.A., 1977. An effective heuristics for the m-tour traveling salesman problem with some side constraints. Operational Research 25, 521524. Russell, R.A., 1995. Hybrid heuristics for the vehicle routing problem with time windows. Transportation Science 29, 156166.

837

Solomon, M.M., 1987. Algorithms for vehicle routing and scheduling problems with time window constraints. Operational Research 35 (2), 254265. Taillard, E., Badeau, P., Gendreau, M., Guertin, F., Potvin, J.-Y., 1996. A new neighborhood structure for the vehicle routing problem with time windows. Transportation Science. Taillard, E., Badeau, P., Gendreau, M., Guertin, F., Potvin, J.-Y., 1997. A tabu search heuristic for the vehicle routing problem with soft time windows. Transportation Science 31, 170186. Thangiah, S.R., 1995. Vehicle routing problems with time windows using genetic algorithms. In: Chambers, L. (Ed.), Practical Handbook of Genetic Algorithms: New Frontiers, Vol. II. CRC Press, Boca Raton, FL, pp. 253277. Thangiah, S.R., 1999. A hybrid genetic algorithm, simulated annealing and tabu search heuristics for the vehicle routing problem with time windows. In: Chambers, L. (Ed.), Practical handbook of Genetic Algorithms: Complex Coding Systems, Vol. III. CRC Press, Boca Raton, FL, pp. 253277. Thangiah, S.R., Nygard, K., 1992. School bus routing using genetic algorithms. Proceedings of the Applications of Articial Intelligence X: Knowledge Based Systems, Orlando. Thangiah, S.R., Petrovic, P., 1998. Introduction to genetic heuristics and vehicle routing problems with complex constraints. In: Woodruff, D.L. (Ed.), Advances in Computational and Stochastic Optimization, Logic Program and Heuristics Search. Kluwer Academic Publishers, Dordrecht, pp. 253280. Thangiah, S.R., Vinayagamoothy, R., Gubbi, A., 1993. Vehicle routing with time deadlines using genetic and local algorithms. Proceedings of the Fifth International Conference on Genetic Algorithms, Morgan Kaufman, NY, pp. 506513. Thangiah, S.R., Osman, I.H., Vinayagaoorthy, R., Sun, T., 1994. Algorithms for vehicle routing problems with time deadlines. American Journal of Mathematical and Management Science 13 (3&4), 323355. Thangiah, S.R., Potvin, J.-Y., Sun, T., 1996. Heuristics approaches to vehicle routing with backhauls and time windows. International Journal of Computers and Operations Research 23 (11), 1043 1057. Thompson, P.M., Psaraftis, H., 1993. Cyclic transfer algorithms for multi-vehicle routing and scheduling problems. Operations Research 41, 935946. Yellow, P., 1970. A computational modication to the saving method of vehicle scheduling. Operational Research Quarterly 21, 281283.