Anda di halaman 1dari 12

Fast local searches for the vehicle routing problem with time windows Olli Braysy INFOR; Nov

2002; 40, 4; A B M N F O R M Global pg. 319

FAST L O C A L S E A R C H E S FOR T H E V E H I C L E ROUTING P R O B L E M W I T H T I M E WINDOWS


OLLI BRYSY SINTEF Applied Mathematics, Department of Optimization, P.O. Box 124 Blindent, N-0314 Oslo, Norway, E-mail: Olli. Braysy@sintef. no ABSTRACT The purpose of this paper is to present new deterministic local searches for solving the vehicle routing problem with time windows. The proposed algorithms are based on a new three-phase approach. In the first phase an initial solution is created with one of the two proposed route construction heuristics. In the second phase a special local search operator based on ejection chains is used to reduce the number of routes. Finally, in the third phase well-known Or-opt exchanges are used to minimize the total distance of the routes. The findings of computational experiments indicate that the proposed methods arc competitive with the best approaches proposed earlier in the literature in terms of solution quality, while being much faster. Moreover, the proposed algorithms may easily be used to create initial solutions for a wide variety of vehicle routing algorithms. RSUM L'objectif de cet article est de prsenter de nouvelles recherches locales dterministes pour rsoutire le problme de tournes de vhicules avec fentres de temps. Les algorithmes proposs sont bass sur une nouvelle approche en trois tapes. Dans la premire tape, une solution de dpart est obtenue avec l'une des deux heuristiques de construction de route proposes. Dans la seconde tape, une recherche spciale base sur des chanes d'jection est utilise de manire rduire le nombre de routes. Finalement, dans une troisime tape, un voisinage classique Or-Opt est utilis pour minimiser la longueur totales des routes. Les rsultats des calculs exprimentaux indiquent que les mthodes proposes sont, en terme de qualit, concurrentielles avec les meilleures approches proposes auparavant dans la littrature tout en tant nettement plus rapides. En outre, les algorithmes proposs peuvent facilement tre utiliss pour obtenir des solutions initiales pour un large ventail d'algorithmes de tournes de vhicules. \. INTRODUCTION

In this paper we focus on the Vehicle Routing Problem w i t h Time Windows ( V R P T W ) , an important problem occurring in many distribution systems. To describe the VRPTW, let G = (C, A) be a directed graph, where C= {c ,c ,...,c } is a vertex set, and/1 = {(c,,c ) : i ^ /') is an arc set. Vertex c denotes a depot and the remaining vertices o f C represent customer locations. Each arc (c,,Cy) has an associated nonnegative distance d- and a nonnegative travel time r,y. The V R P T W consists o f designing a set o f least cost vehicle routes such that:
0 t n 7 0

a. Every route starts and ends at the depot, c . b. Every customer o f C (excluding the depot) is visited exactly once by exactly one vehicle. c. The total demand o f any vehicle route does not exceed the vehicle capacity.
{)

d. The service o f each customer c, begins within a service time window (<?,, /,) where e- is the earliest time that service can begin and /, is the latest time that service can begin.
t

Each customer c,- has a required service time ,v , and arrival at customer c before time <, requires ? a wait time o f w- The V R P T W has multiple objectives in that the goal is to minimize not only the number of vehicles required, but also the total travel time, and total travel distance incurred by the fleet o f vehicles. In this paper we use a hierarchical objective function, where the number o f routes
( i r

Recti. Feb. 2002, Rev. July 2002, Acc. Aug. 2002

INFOR vol. 40, no. 4 Nov. 2002

319

R e p r o d u c e d with permission of the copyright owner.

Further reproduction prohibited without permission.

320

O. B R A Y S Y

is the primary objective, followed by the minimization o f the total traveled distance. Some o f the most useful applications o f the V R P T W include bank deliveries, postal deliveries, industrial refuse collection, national franchise restaurant services, school bus routing, and security patrol services. Because o f its high practical importance the V R P T W has been the focus o f intensive research efforts for both optimization and heuristic approaches. Special purpose surveys for the V R P T W can be found in Golden and Assad (1986), Desrochers ct al. (1988), Golden and Assad (1988), Solomon and Desrosiers (1988), Dcsrosiers et al. (1995), Cordeau et al. (2001) and Braysy and Gendreau (2001a and 2001b). Because o f the high complexity o f the V R P T W , the majority o f research has focused on heuristics. These heuristics can be divided into three main categories: construction heuristics, improvement heuristics (local searches) and metaheuristics. A number o f route construction algorithms are proposed in Solomon (1987). These construction algorithms select and insert nodes sequentially until a feasible solution has been created. Nodes are chosen based on a specific cost minimization criterion subject to the restriction that the insertion does not create a violation o f vehicle capacity or time window constraints. Different variants o f construction heuristics can be found also in Potvin and Rousseau (1993), Bramcl and Simchi-Levi (1996), Dullaert (2000) and Ioannou et al. (2001). For more details on these approaches, we refer to Braysy and Gendreau (2001a) Most o f the recently published V R P T W heuristics use a two-phase approach. First, a construction heuristic is used to generate a feasible initial solution. During the second phase, an iterative improvement heuristic is applied to the initial solution. These route improvement methods iteratively modify the current solution by performing local searches for better neighboring solutions. Generally, a neighborhood comprises the set o f solutions that can be reached from the present one by swapping a subset o f k arcs between solutions. For the most successful applications to the V R P T W , see Thompson and Psaraftis (1993), Potvin and Rousseau (1995), Russell (1995), Shaw (1997 and 1998), Cordone and Wolfler-Calvo (1998) and Caseau and Laburthe (1999). For more details, we refer to Braysy and Gendreau (2001a). To escape from local optima, the improvement procedure can be embedded in a metahcuristic such as simulated annealing, tabu search or genetic algorithm. Genetic algorithms belong to the classical local search framework where improvement is sought with each move in the neighborhood o f the current solution. In contrast, tabu search and simulated annealing are part o f a new paradigm that allows the selection o f worse solutions once a local optimum has been reached. For most successful recent applications, see Rochat and Taillard (1995), Taillard et al. (1997), Gambardella et al. (1999), Homberger and Gehring (1999), Cordeau et al. (2001), Gehring and Flomberger (2001), Braysy (2001b), Berger et al. (2001) and a recent survey by Braysy and Gendreau (2001b). In the past, the focus in the routing research has been mainly on improving solution quality, regarding computation speed as secondary. Nevertheless, there is an industry need for fast, highquality solution methods for routing problems. Because o f the integration o f routing problems in other company activities, the corresponding extended routing problems require faster solution methods to obtain results in an acceptable time span. Especially for dynamic routing problems, where decisions on load acceptance and routing must be made in real-time, the computational speed o f a solution approach is crucial. The main contribution o f this paper is the development of several fast local search strategies for the VRPTW. The proposed algorithms are shown to be very efficient approaches for solving the vehicle routing problem w i t h time windows, producing results that outperform recent local search approaches. They arc also competitive with the best metaheuristics reported in the literature in terms o f solution quality, while being much faster. Flexible, the local search algorithms presented in this paper can also be combined with different routing algorithms to further improve solution quality. The remainder o f this paper is organized as follows. A n overview o f the solution strategy is first given, and then, the different components o f the algorithms are described in Section 2. First, the construction heuristics used to create initial solutions are described. Second, the basic local

R e p r o d u c e d with permission of the copyright owner.

Further reproduction prohibited without permission.

VEHICLE ROUTING PROBLEM WITH TIME WINDOWS

321

search techniques used in computing new solutions are depicted. Then, the ejection chain method is introduced. The computational experiments assessing the value o f the proposed approach are presented in Section 3. Also, a comparative performance analysis involving various local searches and metaheuristics is briefly reported. Section 4 concludes the paper. 2. T H E P R O B L E M S O L V I N G M E T H O D O L O G Y 2.1. Overview We suggest a new three-phase approach for solving vehicle routing problems. In the first phase an initial solution is created using one o f the two proposed construction heuristics. In the second phase an effort is made to reduce the number o f routes using an ejection chain-based approach. The first two phases are repeated for different parameter settings defined for the construction heuristics. In the third phase, well-known Or-opt exchanges (Or, 1976) are used to minimize the total distance. The basic idea in Or-opt exchanges is to reinsert segments o f consecutive customers to other feasible locations, and accept the move i f the insertion improves solution value. The Or-opt is applied only to solutions having the smallest number o f routes, as found during the first two phases. A t the start, the routes o f each solution are reordered such that the routes having some bad features or weaknesses are considered first for improvement. Here we consider long waiting time and long average distance w i t h respect to the number o f customers on the route as bad features. Parameter p is used to control the weight o f these two factors. The larger the value o f [3 the more long average traveled distance is emphasized in the reordering o f the routes. To speed up the feasibility checks on each insertion or move, we used push-forward and push-backward strategies, as introduced in Solomon et al. (1988). The proposed algorithm is as follows: Step 1. Use H y b r i d Construction or Merge heuristic to create an initial solution. Step 2. Repeat the route elimination procedure until no more routes can be eliminated. Step 3. Repeat steps 1 and 2 using all the parameter values w i t h i n the specified limits. Store the created solutions. Step 4. Step 5. Step 6. Step 7. Identify all the created solutions w i t h the smallest number of routes and insert them into set RB. Reorder the routes in solution 5,- according to parameter (3 and improve 5,- using the Or-opt procedure. Repeat step 5 for all S in RB and update the best solution found, S , i f needed. Return S .
: h h

2.2. Route Construction and L o c a l Search Heuristics T w o construction heuristics are proposed to create the initial solutions, namely Hybrid Construction heuristic ( H C ) and Merge Heuristic ( M H ) . They are described below, followed by a description o f five local search procedures B 1 , B 2 , B3, B4 and B5 that are based on HC and M H heuristics, ejection chain-based route elimination procedure and Or-opt exchanges. 2.2.1. Hybrid Construction Heuristic The hybrid construction heuristic is a cheapest-insertion-based heuristic introduced in Braysy (2001b). It borrows its basic ideas from the studies o f Solomon (1987) and Russell (1995). Routes are built one at a time in a sequential fashion and after k customers have been inserted in the current partial route, the route is reordered using Or-opt exchanges. Here /< is a constant parameter value decided by the user. The routes are initialized by selecting the first initialization (seed) customer (initializing the first route) among the 8 customers farthest away from the depot and from each other. The subsequent seeds are selected among the 30% o f the customers farthest away from the depot in clockwise or counterclockwise sweep, or such that the next seed selected is the closest unrouted customer to the previously selected seeds. For more details, we refer to Braysy (2001a).

R e p r o d u c e d with permission of the copyright owner.

Further reproduction prohibited without permission.

322

O. B R A Y S Y

Once the first customer is selected, the unrouted customers are examined one by one and the customer that minimizes the weighted combination o f additional detour and waiting time is selected and inserted in the best feasible insertion place. Here we do not consider all customers for insertion. Instead, we consider only customers that are geographically close to at least one o f the previously inserted customers on the route. The closeness d is a fixed constant determined by the user. The customers farthest away from the depot are usually the most difficult ones to route, since there are often only a few feasible insertion places available for them. Therefore, we favor the selection o f these customers by subtracting from the insertion cost the distance o f the corresponding customer to the depot multiplied by a user defined parameter a . M o r e formally
3

the cost function for customer c is


u
C

u = \ *u

+ 2 ' ^ - 3 *4)>
a

(1)

where
D
u

=d
u

iu

+ d j-d
U

ip

(2)

w = w;-wt,
a i + a Notations d , d
ju 2

O)

= 1, a

> 0.

uJ

and

refer to the distance between the corresponding pair o f customers (c,-,


(

c ), ( f , c ) and (c
H H ;

Cj) and W* and W ", correspond to the total waiting time before and after the
0ll

insertion respectively. Finally d


3

is the distance from the customer c to the depot and a,,


u

and ct are parameter values determined by the user. For more details, we refer to Briiysy (2001a and 2001b). In the experimental part the combination o f HC heuristic, route elimination procedure and Or-opt exchanges is marked by B l . 2.2.2. Merge Heuristic

The merge heuristic draws its basic concepts from the well-known savings heuristic o f Clarke and Wright (1964). We implemented a parallel version o f the savings heuristic, as in Clarke and Wright (1964). We begin by serving each customer by a separate route and then construct the solution by repeatedly merging a pair o f routes producing the greatest saving as long as the time or capacity constraints are not violated. The saving 5,-- incurred by combining routes i and j is calculated using the weighted combination o f distance and waiting time to evaluate the total cost o f each individual route k, TC = a, TD + a TW ,
k k 2 k

where TD and TW are the total distance


k k ri

and total waiting lime on route k. Then, the cost o f the merged route {TC )
t

is deducted from the

j o i n t cost o f routes / a n d i . e . , S- = TC + TC- - TC-. Moreover, an attempt is made to improve the combined route by trying to reorder customers, to reduce TC^, before evaluating the saving incurred by uniting the two routes. The improvement is tried when m more customers have been added to the merged route, using the intra-route improvement heuristic O-opt, described in Briiysy (2001b). The basic idea is to rebuild routes with the cheapest insertion heuristic by using more than one customer to initialize the routes. More precisely, instead o f selecting just one seed customer, a user-defined number, /, o f customers are selected from the current route such that they are geographically as dispersed as possible. These seed customers are then put in all feasible orders and each o f the partial routes initialized by these / customers is rebuilt with the hybrid construction heuristic by inserting customers in increasing order of their time window w i d t h . After k customers have been inserted, the route is reordered using the Or-opt operator. To avoid double work, the previously calculated savings are stored into a matrix. In the experimental part, the combination o f M H , route elimination procedure and Or-opt exchanges is marked by B2. In addition to testing B l and B2 heuristics independently, they were combined simply by running both methods and taking the best result. In the experimental part we refer to this method by notation B3. According to our experience most approaches developed for the V R P T W have difficulties in obtaining good output regarding the number o f routes. To alleviate this problem we suggest an implementation that is solely designed to create quickly initial solutions with a small

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

VEHICLE ROUTING PROBLEM WITH TIME WINDOWS

323

number of routes. This approach can be combined for example with any metaheuristic or large neighborhood search procedure to minimize the total traveled distance. The basic idea is the same as that used in the implementation o f B 1. However, here we use the knowledge o f the best solutions reported in the literature and the search is stopped once the number o f routes is equal to the best known. Also, the number o f parameter combinations tried is greater than in the B1 implementation. I n the experimental part we use notation B4 for this method. The usage of the best known number of routes can be justified by the fact that in practice the number of vehicles is often known, i.e., the company has a limited number o f vehicles (we assume that each route is served by a separate vehicle). However, when the number o f routes/vehicles is not known in advance, there are also several ways to handle the problem algorithmically. For example one can use some other heuristic or metaheuristic method to estimate the number of routes or alternatively use different lower bounds such as solutions from different B i n Packing Problems ( K o n toravdis et al., 1995). One possibility is also to use the information gathered during the search, i.e., since we use very powerful ejection chains to reduce the number o f routes, we often get solutions where the number o f routes is very close to the optimum. Thus, this information can also be used to estimate the number o f routes. Finally, we created a fifth implementation B5 by examining all parameter values specified for B4, without prematurely stopping the search, i.e., without using the knowledge of the best-known solutions. 2.3. Route Elimination Procedure After creating an initial solution, the ejection chain-based technique introduced in Briiysy (2001b) is used to reduce the number o f routes. The basic idea in ejection chains is to combine a series o f simple moves into a compound move. I n a V R P context, these simple moves refer to removal o f a customer from its route and reinsertion o f the removed customer in another route. The goal is to "make room" for a new customer in a route by first removing another customer from the same route. In each phase w i t h i n the ejection chain, one customer remains unrouted. The removal and insertion procedures are repeated until one can insert a customer to another route without the need to remove (eject) any customer. For more details on ejection chains, and their application to V R P and V R P T W we refer to Glover (1991, 1992), Rego (1998, 2001) Rousseau et al. (2000) and Caseau and Laburthe (1999), respectively. The main innovation in the ejection chain approach described in Braysy (2001b) is the reordering o f the routes within the ejection chain. First, the customer c, originally served by route r is inserted at a location on another route r, that least increases the value o f cost function ( 1 ) with a, = 0.5 and a = 0. I f the time w i n d o w constraints are violated, simple intra-route rinsertions are tried to make r, feasible after inserting c . The basic idea of these reinsertions is to reduce lateness in the infeasible route by trying to serve some customers in alternate locations within the same route. We used the first-accept strategy, i.e., accept every move that reduces lateness. A l l routes are considered for elimination, one at a time, starting from the routes having the smallest number o f customers. I f direct relocation o f a customer c, on the route r currently considered for elimination to an alternate route is not possible, the ejection chain procedure with breadth-first strategy is initialized using c I n the breadth-first search, we try first all possible chains involving one ejection and two insertions, then chains w i t h two ejections and three insertions and so on. I f we are not able to relocate all customers on r , a special post-processing is applied. Here we insert as many customers as possible from the neighboring routes to the just examined route r to facilitate eliminating some other route. Since we must keep route r , it is reasonable to utilize its resources as w e l l as possible. In addition, several strategies/restrictions were used to speed up the algorithm:
e 3 ; e r e e e

- The maximum length o f ejection chain and number o f w i t h i n route reinsertions are restricted to c and n, respectively. - To restrict the number o f chains to explore, we set a limit 1 to the allowed increase in distance.

Reproduced with permission of the copyright owner.

Further reproduction prohibited without permission.

324

O. B R A Y S Y

Thus, i f a certain insertion causes the total distance o f the target route to become 7 times longer, the corresponding insertion and/or chain is ignored. We used push-forward and push backward strategies introduced in Solomon et al. (1988) to speed up the feasibility checks of each insertion or reinsertion within the reordering operator. In addition, to further increase the efficiency o f feasibility checks, we maintained in memory the arrival times and the latest possible arrival times at each customer. - To minimize the number o f insertions tried, we try to eliminate first routes having the smallest number o f time-constrained customers, and we first try to insert the customers in each phase to the routes that are geographically closest. The geographical closeness is defined as the smallest distance between the currently considered customer c and any customer on the destination route. Moreover, since it is computationally prohibitive to try to eliminate a long route, only the shortest route is considered for elimination in case o f problems with over 15 customers per route.
1 (

I f the distance between the customer c currently considered for insertion and all the customers in a given route r, exceeds a user-defined l i m i t d, the insertion o f customer c, is not even tried in r .
; t

3. C O M P U T A T I O N A L E X P E R I M E N T 3 . 1 . Problem Data and Parameter Values To analyze the performance o f our algorithm, an experiment was conducted over 56 VRPrW benchmarks, partitioned in six data sets denoted R l , C I , R C I , R2, C2 and RC2 (Solomon, 1987). The problems vary in fleet size, vehicle capacity, travel time and spatial and temporal customer distribution (position, time w i n d o w density, time window width, and service time). Each problem instance involves a central depot, 100 customers to be serviced, as w e l l as constraints imposed on vehicle capacity and customer visit or delivery time (time windows). C I and C2 data sets are characterized by a clustered customer distribution, whereas R1 and R2 refer to randomly distributed customers. Data sets R C I and RC2 represent a combination o f random and clustered customer distributions. The travel time between two locations corresponds to the Euclidean distance. The algorithms were implemented in JAVA and the computational experiments were conducted using a Pentium 200 M H z computer. We found it computationally intractable to optimize the value for each parameter separately. Therefore, we tuned the parameter values only once by selecting first a parameter setting based on intuition. Then, a few (310) intuitively selected values were tried for each parameter, while keeping the other parameter values fixed. To reduce the workload, only a set o f four problems were used in the tuning. Each time, we selected a parameter value that gave the best average output for the selected four test problems. The general parameters used by all our algorithms remained fixed during the computations. However, in some cases the parameter values depend on the characteristics o f the problem. We separate two groups o f problems, based on the average number o f customers on each route to control the complexity o f the search. M o r e precisely, we include problem sets R l , C I and R C I in group I and the other problem sets R2, C2 and RC2 in group I I . The parameters and their values are: d = 20 and c = 9 in group I , d = 30 and c = 4 in group I I , h = 5, (3 = 0.5, k = 3 , 1 = 1.15, i = 4, m = 7, a,: 0.7-1.0 (in increments o f 0.05 and 0.1 units for B2 and for B l , B 4 and B5, respectively), a : 0.5-1.7 (in increments o f 0.2 units). In case o f B l we used the following values forot : 0.5, 1.1, 1.7 .
3 2 3

1 Wc define a customer as time-constrained if the width of the time window is less than 50% of the depot's time window. 2 The values were selected by taking the lower and upper bound and the middle value of the range defined for B4 and B5.

R e p r o d u c e d with permission of the copyright owner.

Further reproduction prohibited without permission.

VEHICLE ROUTING PROBLEM WITH TIME WINDOWS

325

In order to keep the memory requirements reasonable, a m a x i m u m l i m i t o f 300 is set to the ejection chains stored in memory. I n case the elimination procedure returns the same number of routes, nf, each time, we compare n{.to the number o f routes in the initial solution, n' . I f the values are the same, we assume that the problem is easy, and we have obtained the m i n i m u m number o f routes. Otherwise, the m a x i m u m length o f ejection chains is increased to c = 10, the geographical closeness range is increased to d = 0.4 and 7 = 1.65, and the elimination procedure is repeated i n order to introduce more power to the search.
r

The reader must note that even i f there are a lot o f different parameters, in general the results are not sensitive to changes in parameter values. I n most cases the effect o f the parameter values on total distance was less than 1%, which made it possible to find a single robust value that is not dependent on the problem. The only exceptions were parameters a,, a and the seed selection schemes that had a clearly bigger impact on the results (in some cases over 50%). We found it impossible to determine a robust value for them that would give good results for all test problems. Therefore, we tried several values for these parameters w i t h i n specified ranges, as depicted above. The parameter sensitivity is discussed in more detail in Braysy (2001a).
3

3.2. Computational Results In Table 1 we compare the results obtained w i t h the proposed five local search procedures with the results o f recent local searches and metaheuristics. We included in Table 1 only approaches where sufficient information is provided by the authors. A t least the computer, number o f computational runs as well as the time consumption, number o f vehicles and total traveled distance for each problem group must be reported to make the comparison possible. The first column to the left gives the authors and columns R l , R2, C I , C2, RC1 and RC2 present the average number o f vehicles and average total distance w i t h respect to the six problem groups of Solomon (1987). The C N V / C T D column indicates the cumulative number of vehicles ( C N V ) and cumulative total distance ( C T D ) over all 56 test problems. The rightmost column describes the computer, number o f runs, and the C P U time used to obtain the reported results. Two CPU time values are described: the one reported by the authors and in the parenthesis the modified C P U time. These modified CPU times are scaled to equal the running times on a Sun Sparc 10/50 using Dongarra's (1998) factors. I f several runs are required to get the reported results, the computational times are multiplied by this number to see the real computational effort. One must note that these modified times are only indicative and should be used only to get some kind o f picture o f the magnitude o f the running times. The Table is divided in three parts, presenting results o f some recent local searches and metaheuristics, and our local searches, respectively. In most earlier papers the results are ranked according to a hierarchical objective function, where the number o f routes is considered as the primary objective, and for the same number o f routes, total traveled distance is considered as the secondary objective. Therefore, a solution requiring fewer routes is always considered better than a solution with more routes, regardless o f the total distance. The rationale behind this approach is that each additional route requires very costly investments in regard to new vehicles and drivers causing fixed costs for a long time period. Thus one assumes that each route is served by a separate vehicle. In real-life problems this assumption does not always hold, and also other objectives are considered instead o f seeking just a solution w i t h the least number o f vehicles. Here we used the hierarchical objective function to make the comparison w i t h previous approaches possible. According to Table 1, it seems that our approaches are very competitive regarding the number of routes. Only Braysy (2001b) reports slightly lower C N V than B 4 and B5. A l l other earlier approaches except Homberger and Gehring (1999) and Gehring and Homberger (2001) report greater C N V . The better results in the above mentioned three papers are, however, obtained at the cost o f clearly higher computational effort, which may make it prohibitive to use them in large-scale or dynamic real-life problems. The reason for the good performance regarding the number o f routes can be found in the robust route elimination procedure. In general, the differ-

R e p r o d u c e d with permission of the copyright owner.

Further reproduction prohibited without permission.

a ^ 5 o
NO

N " <N

4S^
< ~ . NO w

s o
00

a
HH-'I

-S
6

oo
O 00 C/3
C 3

>o E c

2 O
C

C N CN ^

p c .

CN

a
S P.

u c
OH CO

cjo S g S
0< OH

a -a p a
H

OH 2 XT

OH

t/3

&

I-H

5 h O CN 5 S m o . - O a ON c P A P
EH ^ H

OH

(H

H H

P u > z u

" F - ON oo <N , - H t - t - ^ t N O N O in N O 0 0 co CN 0 0 C N CN C N c o < N C N CN 0 0 CN C N r H r H r- r H O N r H oo T f OO t ON <t NO 00 CO CO OO ii 00 00 C-NO ON C-r<o m CO co co m CO *n

95

2P?2b^^^^<^oocoir)co>r)^roooi/}r-

u
<N

OH

C ^ ^ r O M f c i d c c i a f O O N C ^ O O f ^ ^ f ^ ^ r O ^ r ^ ^ O N c N c o c o o N x t ^ t - o Co C N >-< ^ H rn _ H ( N

x m O N O ^ X [ ^ o o r H f f ) ^ o o ^ M m t ^

f O M C f ) 0 \ O t ^ t ^ o o n ^ - S \ c n > O i o O O 0 \ c e o < ] rH H ^ - r t O H O H ^ O H o O H C h H ^ H C O H ^

>o

co

co

co

CO O r O O C~-

> i T - H o co co co \ D O ON OO O O ^roi^rf^dc^^codndnoSco co i ci r - i ^H O N O N O N O N O O
CO C O I O CO C O ON O N

CO

CO

ON 0 0 < 0 1 0 " 0 O O O O N 0 0 \ 0 co co co o ^ t ; ^ t r o C N r o o co oo N t d o o d c o ' d c d d ^ d o d d d C N r H C OO r H C oN r t c oN r H C oN r H rt C O o o o o oN H e n 00

CN

Pi

rH O ON OO ON O N O N O N N O O C O C O C N r - - ^ C N O O ON NO "^f O N O N c" C O C O l ^ ; O N 0 0 x t C O 0 0 C O o CN C O NO C O O N i o n \ t n c n c s o i c N | ' 0 , r i d c i d NO ON ON C O r - l NO NO < CO ON ON O ON ON

0<

N O r - C O N O C N x J - O O N O O C N C O C O C N N O C N o O O O c O t ^ C ^ N ! ^ C X 5 x t ; T f c O I / ^ c i n x f C O C O O N O r t O N f O O O r H i o 2 c n c i " M ^ d i H ' c i ^ r i d c N ' ' o \ r H HMHWH'tHOMHCN'Hrlrl HHHTf


CO C N C N >-< C N C N C N C

^
^ ^
ON

ON ON ON
ON ON ON
S

>o ON
ON
1

o o CN
N^

ii
'

ON ON

'i
'

,cd :
-fcH
CD

0\
V

'

,a; u t/i

as
<U

p
CL>

t/i

<

c o "S O

CO

o Pi

o E C

CD

Reproduced with permission of the copyright owner.

Further reproduction prohibited without permission.

aa
o .

2
o

x a x '3 x
M

'e

< N

a
3
- H

m >-H o 3 3 c

NO d

888s 8B8B8^ ^ ON f g CN S i - i a a 3 m s j 3 - 3 - 3 - 3
^ N O ^ N O < ^
r

a a -a a c
CD PH
H

-a es a c
4J CO P H I P H

PH

i n P H c o CM

s aCM -PH s

CD

3 H

cg
H

u >
Z

CD

NO o

Tf rr~ >o

Tt NO

H rH

NO

t--

rH o r-< o

CO HH

Tt Tt O l Tt C O Tt roO oO io N N

CO 1 CO H

Ol
ON ON ON

CN

TT

t - NO CO NO o > o o CO o ON T T CN
ON

t-~
11

ON

CO

CO

NO

CO

u
N m N m N ^ N r t f N ] q w q N t N c N f N ] 0 \
fi
N'

C O O O C f J C ^ V l O O l f J C O I n O N W C N l l n ^ O I O C O C O t O
C N C O T t i O l O N O T r r O

C ^ M f ^ O N f c N i c ^ Q O ^ d c c o N C ^ N t C ' i c N i c c l d

p>

iIiIiI

C N CN C N CN

CN

Tt O O O O N T t Cf) C O I/) r-; I f i l n l n i H ^ N t M O O G O C N j O h O O CO r i ci H K H d r i c n H i n H i - H C O C N i d H O rt \ rtM H M H O \ H M H r t H o | r t O H TtTt Tt

o o o o c o m r - o o N o o o < o o o Tt Tt

CO

C O

C O

o CN

o
CO

NO OO ON GO

q
CO

CO CO ON CO

NO

o
CO

00
ON

NO

co

oc
CO

q o o c N C O O
CO

CO ON CO ON CO

O
CO

ON

o
CO

ON

o
CO

OO

OO

o Tt o Tt o
CO ON

CO

in

CO ON CO

< o
ON CO

NO ON CO

CO

o o d

o CO o od
OO

NO

Ol
CO

CO o o CO d CN
GO

o CO o 00 d cN
GO OO

0 o o Tf o 0 N o O CO o C Tt o 3 00 d N O d ci d O l c H C O T^H CO GO GO 00

o CN d CO ^ H
GO OO OO

CO o CN d
ON NO

NO

00

CN CO
OO

C O CN C O CN CN

ON

o
ON

CN

rH CN O N cN

CO

Tt
CO ON

NO ON

m CN N O C O N O C O N O C O O N CN o o CO CO p- CO 0 0 C O o 00 t-; 0 0 CN N O CN CN C O CN O N ci d ci N O iH rn CO m, CN rCO

cd

< N C N r - 0 0 T t r - O l f ) N 0 i r ) c o l > T t O r ocHoinOHHtcSNtfjiciHiNjomoh H c s c N " r ^ N ^ c i - i N ^ N o o c i r i c ^ c d c i c > iICN^Ci-Hi-H^iIIOTliOrHint|inrHCTN*-HCO C N C N C N C N C N C N C N C N C N

t al. (21

(2001)

Li et

Berg

B3

B2

Tt m

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

B5

C D I H C D

paper)

paper)

paper)

paper)

L E

paper)

o o CI

328

O. B R A Y S Y

ence in cumulative number o f vehicles between the approaches in Table 1 is about 6%, which is in our opinion remarkable, considering the huge fixed costs caused by buying an additional vehicle and hiring a new driver for it. As one can observe from Table 1, all of our five approaches outperform the previous local searches in terms o f solution quality. As for the melaheuristics, only Homberger and Gehring (1999), Gehring and Homberger (2001) and Braysy (2001b) report better results, while being multiple times slower. The slight overestimation o f total distance with respect to the best values reported in the literature can be justified by the fact that our approaches arc much faster than previous approaches reporting equal CNV, especially i f one considers that our algorithms are implemented using JAVA, which is according to our experience much slower than C-programming language . The main advantage o f using JAVA is the clearly increased development speed due to automatic memory handling mechanisms. This makes it possible to quickly code and test new ideas. Here one must note that due to the conflicting nature o f route number and distance optimization, the comparison over total distance values is reasonable only for the same number o f routes. We consider the given differences in time consumption significant especially in dynamic and very large-scale problems, where the speed o f the solution method is o f prime concern. Moreover, as our local searches are deterministic and simple, they can easily be used to create initial solutions for more complex approaches to further improve the solution quality, i f more time is available. Especially B4 was designed only to quickly create initial solutions with a small number o f vehicles, and to be hybridized w i t h some metaheuristic for distance minimization. According to Table 1, B4 achieves its objective very well by giving solutions with the lowest k n o w n number o f vehicles to all Solomon's benchmark problems, except one, while being more than 18 times faster than any previous approach. Local searches by Russell (1995) and Antes and Derigs (1995) seem to be faster than our approaches, but they fall clearly behind in solution quality.
3

4. C O N C L U S I O N S We proposed five variants o f deterministic local searches that use a new three-phase approach. In the first phase an initial solution is created using one o f the two proposed construction heuristics. Then, a special route-elimination operator, based on a new type o f ejection chains is used to minimize the number o f routes. Finally, in the third phase, the created solutions are improved in terms o f distance using well-known Or-opt exchanges. The computational testing o f the proposed methods was carried out on the 56 test problems o f Solomon (1987). The proposed local search methods were compared with the best previous local searches and metaheuristics found in the literature. The results show that the methods proposed in this paper arc very efficient approaches for solving the vehicle routing problems with time windows, producing results that outperform recent local search approaches and are competitive w i t h the best metaheuristics in terms o f solution quality, while being much faster. Our local searches show good performance especially regarding the number o f routes that is often considered to be the primary objective. The fact that only Braysy (2001b) is able to report lower cumulative number of vehicles over all test problems questions the performance o f metaheuristics in optimizing the number o f routes. On the other hand, for distance minimization purposes, metaheuristics seem to be necessary to avoid being trapped in local minima. A s heuristics need to be especially effective for dynamic and very large-scale problems, we consider our approaches practically relevant. Future research w i l l be conducted to evaluate the performance o f different combinations w i t h metaheuristics to further improve the solution quality.

3 We implemented a modification of the approach described in this paper using both JAVA and C++, and found that JAVA is approximately 10 times slower than C++ on the average. This is not taken into consideration in running times in Table 1.

R e p r o d u c e d with permission of the copyright owner.

Further reproduction prohibited without permission.

VEHICLE ROUTING PROBLEM WITH TIME WINDOWS

329

ACKNOWLEDGEMENTS This work was partially supported by the E m i l Aaltonen Foundation, Liikesivistysrahasto Foundation and the TOP program funded by the Research Council o f Norway. This support is gratefully acknowledged. REFERENCES Berger, J., M . Barkaoui and O. Braysy (2001), "A Parallel Hybrid Genetic Algorithm for the Vehicle Routing Problem with Time Windows", Working Paper, Defence Research Establishment Valcartier, Canada. Bramel, J., and D. Simchi-Lcvi (1996), "Probabilistic Analyses and Practical Algorithms for the Vehicle Routing Problem with Time Windows", Operations Research 44, 501-509. Braysy, O. (2001a), "Local Search and Variable Neighborhood Search Algorithms for the Vehicle Routing Problem with Time Windows", Doctoral Dissertation, Department of Mathematics and Statistics, University of Vaasa, Finland. Braysy, O. (2001b), "A Reactive Variable Neighborhood Search Algorithm for the Vehicle Routing Problem with Time Windows", to appear in INFORMS Journal on Computing. Braysy, O. and M . Gendreau (2001a), "Route Construction and Local Search Algorithms for the Vehicle Routing Problem with Time Windows", Internal Report STF42 AO 1024, SINTEF Applied Mathematics, Department of Optimisation, Norway. Braysy, O. and M . Gendreau (2001b), "Metaheuristies for the Vehicle Routing Problem with Time Windows", Internal Report STF42 A01025, SINTEF Applied Mathematics, Department of Optimisation, Norway. Cascau, Y. and F. Laburthe (1999), "Heuristics for Large Constrained Vehicle Routing Problems", Journal of Heuristics 5, 281-303. 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., 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. Cordeau, J.-F., G. Desaulniers, J. Desrosiers, M . M . Solomon and F. Soumis (2001), "The VRP with Time Windows", In The Vehicle Routing Problem, SIAM Monographs on Discrete Mathematics and Applications, P. Toth and D. Vigo (eds.). SIAM, Philadelphia, 157-194. Cordone, R. and R. Wolfler-Calvo (2001), "A Heuristic for the Vehicle Routing Problem with Time Windows", Journal of Heuristics 7, 107-129. Desrochers, M., J.K. Lenstra, M.W.P. Savclsbergh and F. Soumis (1988), "Vehicle Routing with Time Windows: Optimization and Approximation", In Vehicle Routing: Methods and Studies, B. Golden and A. Assad (eds.). Elsevier Science Publishers, Amsterdam, 65-84. Desrosiers, J., Y. Dumas, M . M . Solomon and F. Soumis (1995), "Time Constrained Routing and Scheduling", In Handbooks in Operations Research and Management Science 8: Network Routing, M.O. Ball, T.L. Magnanti, C.L. Monma and G.L. Nemhauser (eds.). Elsevier Science Publishers, Amsterdam, 35 139. Dongarra, J. (1998), "Performance of Various Computers Using Standard Linear Equations Software", Report CS-89-85. Department of Computer Science, University of Tennessee, U.S.A. Dullaert, W. (2000), "Impact of Relative Route Length on the Choice of Time Insertion Criteria for Insertion Heuristics for the Vehicle Routing Problem with Time Windows", in Proceedings of the Rome Jubilee 2000 Conference Improving Knowledge and Tools for Transportation and Logistics Development: 8th Meeting of the Euro Working Group Transportation, B. Maurizio (ed.). Faculty of Engineering, University of Rome, Rome, 153-156. Gambardella, L.M., E. Taillard and G. Aga/.zi (1999), "MACS-VRPTW: A Multiple Ant Colony System for Vehicle Routing Problems with Time Windows", In New Ideas in Optimization, D. Come, M. Dorigo and F. Glover (eds.). McGraw-Hill, London, 63-76. Gehring, FI. and J. Homberger (1999), "A Parallel Hybrid Evolutionary Metaheuristic for the Vehicle Routing Problem with Time Windows", In Proceedings of EUROGEN99 - Short Course on Evolutionary Algorithms in Engineering and Computer Science, Reports of the Department of Mathematical Information Technology, Series A. Collections, No. A 2/1999, K. Miettinen, M . Miikcla and J. Toivancn (eds.). University of Jyvaskylii, Jy vaskylii, 57-64. Gehring, H. and J. Homberger (2001), "Parallelization of a Two-Phase Metaheuristic for Routing Problems with Time Windows", Asia-Pacific Journal of Operational Research 18, 35^17. Glover, F. (1991), "Multilevel Tabu Search and Embedded Search Neighborhoods for the Traveling Salesman Problem", Working Paper. College of Business & Administration, University of Colorado, U.S.A. Glover, F. (1992), "New Ejection Chain and Alternating Path Methods for Traveling Salesman Problems",

R e p r o d u c e d with permission of the copyright owner. Further reproduction prohibited without permission.

330

O. B R A Y S Y

In Computer Science and Operations Research: New Developments in Their Interfaces, O. Balci, R. Sharda, and S. Zenios (eds.). Pergamon Press, Oxford, 449-509. Golden, B.L. and A.A. Assad (1986), "Perspectives on Vehicle Routing: Exciting New Developments", Operations Research 34, 803-809. Golden, B.L. and A.A. Assad (1988), Vehicle Routing: Methods and Studies. Elsevier Science Publishers, Amsterdam. Homberger, J. and H. Gchring (1999), "Two Evolutionary Metaheuristics for the Vehicle Routing Problem with Time Windows", IN FOR 37, 297-318. Ioannou, G., M. Kritikos and G. Prastacos (2001), "A Greedy Look-Ahead Heuristic for the Vehicle Routing Problem with Time Windows", Journal of the Operational Research Society 52, 523-537. Kontoravdis, G.A. and J.F. Bard (1995), "A GRASP for the Vehicle Routing Problem with Time Windows", INFORMS Journal on Computing 7, 10-23. Or, I . (1976), "Traveling Salesman-Type Combinatorial Problems and their Relation to the Logistics of Regional Blood Banking", Ph.D. Thesis. Department of Industrial Engineering and Management Sciences, Northwestern University, Evanston, U.S.A. Potvin, J.-Y. and J.-M. Rousseau (1993), "A Parallel Route Building Algorithm for the Vehicle Routing and Scheduling Problem with Time Windows", European Journal of Operational. Research 66, 331-340. Potvin, J.-Y. and J.-M. Rousseau (1995), "An Exchange Heuristic for Routeing Problems with Time Windows", Journal of Operational Research Society 46, 1433-1446. Rego, C. (1998), "A Subpath Ejection Method for the Vehicle Routing Problem", Management S cience 44, 1447-1459. Rego, C. (2001), "Node Ejection Chains for the Vehicle Routing Problem: Sequential and Parallel Algorithms", Parallel Computing, 27 201-222. Rochat, Y. and E. Taillard (1995), "Probabilistic Diversification and Intensification in Local Search for Vehicle Routing", Journal of Heuristics 1, 147-167. Rousseau, L.-M., M . Gendreau and G. Pesant (2002), "Using Constraint-Based Operators to Solve the Vehicle Routing Problem with Time Windows", Journal of Heuristics 8, 43-58. Russell, R.A. (1995), "Hybrid Heuristics for the Vehicle Routing Problem with Time Windows", Transportation Science 29, 156-166. Shaw, P. (1997), "A New Local Search Algorithm Providing High Quality Solutions to Vehicle Routing Problems", Working Paper. Department of Computer Science, University of Strathclyde, Glasgow, Scotland. Shaw, P. (1998), "Using Constraint Programming and Local Search Methods to Solve Vehicle Routing Problems", In Principles and Practice of Constraint Programming - CP98, Lecture Notes in Computer Science, M . Maher and J.-F. Puget (eds.). Springer-Verlag, New York, 417-^4-31. Solomon, M . M . (1987), "Algorithms for the Vehicle Routing and Scheduling Problems with Time Window Constraints", Operations Research 35, 254-265. Solomon, M.M., E.K. Baker and J.R. Schaffer (1988), "Vehicle Routing and Scheduling Problems with Time Window Constraints: Efficient Implementations of Solution Improvement Procedures", In Vehicle Routing: Methods and Studies, B. Golden and A. Assad (eds.). Elsevier Science Publishers, Amsterdam 85-106. Solomon, M . M . and J. Desrosiers (1988), "Time Window Constrained Routing and Scheduling Problems", Transportation Science 22, 1-13. Taillard, E., P. Badeau, M . Gendreau, F. Guertin, and J.-Y. Potvin (1997), "A Tabu Search Heuristic for the Vehicle Routing Problem with Soft Time Windows", Transportation Science 31,170-186. Thompson, P.M. and H.N. Psaraftis (1993), "Cyclic Transfer Algorithms for Multivehicle Routing and Scheduling Problems", Operations Research 41, 935-946.

Olli Braysy is currently a Research Scientist at SINTEF Applied Mathematics, Department of Optimization in Oslo, Norway. He received the master, licentiate and doctoral degrees from Department of Mathematics and Statistics at the University of Vaasa, Finland in years 1998-2001. His research interests are focused on different routing and scheduling problems and heuristic and mefaheuristic solution methods.

R e p r o d u c e d with permission of the copyright owner. Further reproduction prohibited without permission.

Anda mungkin juga menyukai