Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

European Journal of Operational Research 169 (2006) 606–622

www.elsevier.com/locate/ejor

Scatter search for the vehicle routing problem


with time windows
Robert A. Russell *, Wen-Chyuan Chiang
Finance and Operations Management Department, The University of Tulsa, 600 South College Avenue, Tulsa, OK 74104, USA

Received 16 July 2003; accepted 24 June 2004


Available online 14 October 2004

Abstract

In this paper we use a scatter search framework to solve the vehicle routing problem with time windows (VRPTW).
Our objective is to achieve effective solutions and to investigate the effects of reference set design parameters pertaining
to size, quality and diversity. Both a common arc method and an optimization-based set covering model are used to
combine vehicle routing solutions. A reactive tabu search metaheuristic and a tabu search with an advanced recovery
feature, together with a set covering procedure are used for solution improvement. Our approach led to a robust solu-
tion method, generating solution quality that is competitive with the current best metaheuristics.
 2004 Elsevier B.V. All rights reserved.

Keywords: Scatter search; Tabu search; Metaheuristic; Set covering; Vehicle routing with time windows

1. Introduction solutions are usually generated from a heuristic


and may or may not be improved, but the best
In this paper we use a scatter search framework ÔbÕ solutions are collected. With diversity, the solu-
to solve the vehicle routing problem with time win- tions may be generated from different heuristics to
dows (VRPTW). Our objective is to achieve effec- explore different regions of the solution space. The
tive solutions and to investigate the effects of main idea behind scatter search is that the com-
reference set design parameters pertaining to size, bined solutions will produce a synergistic effect
quality and diversity. Scatter search is a metaheu- and capture information not contained in the indi-
ristic approach that combines solutions from a ref- vidual solutions. In implementation, auxiliary heu-
erence set to achieve improved solutions. The ristic solution methods will be used to evaluate the
reference set is usually generated with quality combinations and generate new solutions. In es-
and diversity in mind. With quality, the initial sence, the newly combined solutions will explore
different parts of the solution space where better
*
Corresponding author. objective function values may possibly be found.

0377-2217/$ - see front matter  2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2004.08.018
R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622 607

Scatter search has been applied successfully to (Rego, 1998). Corberan et al. (2002) developed
varieties of problem types including quadratic an efficient scatter search procedure to solve a
assignment (Cung et al., 1997), flow shop schedul- multiple objective school bus routing problem
ing (Jain and Meeran, 2002), linear ordering in a rural area. The search solution procedure
(Campos et al., 1999), bipartite drawing (Laguna follows the framework of scatter search—con-
and Marti, 1999), graph coloring (Hamiez and structing, improving and combining solutions.
Hao, 2002), and diverse mixed integer solution Experimental testing with real data is used. Grei-
generating (Glover et al., 1999). One most recent storfer (2003) applied a tabu search procedure
paper worthy of attention is Campos et al. (in that makes use of the scatter search paradigm
press) which successfully generalized a scatter for solving a capacitated arc routing problem in
search framework to solve context-independent which the objective is to determine a least-cost
permutation problems, including the band reduc- schedule of routes. A fleet of vehicles located at
tion problem, the linear ordering problem, the a single depot is used to service the route de-
traveling salesman problem, and the single ma- mands. The computational results are comparable
chine sequencing problem. In this framework, ten to other arc routing heuristics.
different types of solution combinations are used. Within the scatter search framework we employ
The scatter search with tabu search improvement metaheuristic methods to guide the search for
yields higher solution quality than commercially effective solutions. The first improvement method
available packages. Readers are advised to refer is a short-term reactive tabu search method that
to Glover et al. (2002) and Glover (1998) for a has been developed for the VRPTW. The search
complete coverage of rich literature in the applica- is short term in that the number of iterations is
tions of scatter search. limited to 10 iterations per customer for a maxi-
Previous research in scatter search for vehicle mum of 500 iterations. The reactive tabu search
routing related problems has been conducted procedure utilizes neighborhood moves, and inten-
by Rochat and Taillard (1995), Rego (2000), sification, and diversification strategies to modify
Corberan et al. (2002), and Greistorfer (2003). routes and escape local optima. This method is
Rochat and Taillard called their approach a context dependent in that it is designed specifically
probabilistic local search although it employed for vehicle routing type problems. We also employ
many of the strategies of scatter search. They gen- an enhanced version of tabu search utilizing an ad-
erated a pool of initial solutions and then proba- vanced recovery strategy that samples from a pool
bilistically chose tours from the pool weighted by of elite solutions. The advanced recovery tabu
their solution quality. Partial tours were created search is employed only at the end of the improve-
and a local search procedure was used to con- ment process since it is more computationally
struct full tours leading to new and possibly bet- intensive than the reactive tabu search.
ter solutions for the ‘‘reference set’’. Rochat and The third improvement technique is an integer
Taillard also used set partitioning as a post opti- programming application of set covering. This
mization technique to combine tours contained in optimization-based method is global in that it
the reference set. They applied their search proce- operates on the set of all routes in the set of en-
dure to both the VRP and VRPTW and obtained hanced or improved solutions to create one elite
very good results including best known solutions solution. It is a ‘‘macro’’ method that operates
a few which are still best known currently. Rego on whole routes and thus differs from the ‘‘micro’’
applied scatter search to the standard vehicle node and edge exchange types of neighborhood
routing problem (VRP) and combined solutions moves that occur between routes. Set covering is
from the reference set through a common arc a relaxation of the set partitioning method and of-
mechanism. He was able to achieve reasonable re- fers one advantage; nodes or customers covered
sults but found that his scatter search solution more than once can be eliminated from one or
quality was not as capable as the best known ap- more routes leading to an improved solution rela-
proaches including his subpath ejection method tive to set partitioning. Set covering is context
608 R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622

independent to the extent that it can be applied to (e) Each customer ci has a service time window
a variety of scheduling problems. (ei, li) where ei is the earliest time that service
In this study we conduct experiments examining can begin and li is the latest time that service
the effects of reference set size and its effect on can begin.
solution quality and computational requirements.
We also explore two different methods of combin- Arrival at customer ci before time ei requires a wait
ing vehicle routing solutions. The primary ap- time of wi and service after time li is infeasible.
proach uses arcs or edges that are common to a Additionally, the total duration of any route (tra-
subset of solutions. A second approach is to com- vel plus service time) can not exceed a preset route
bine sets of routes using the set covering model. duration bound.
We end the experiments with a discussion of the The VRPTW has multiple objectives in that the
effectiveness of the scatter search implementation goal is to minimize not only the number of vehicles
in solving the VRPTW. required, but also the total travel distance incurred
by the fleet of vehicles. Although the priorities can
vary by application, i.e. sometimes total travel
2. The vehicle routing problem with time time is also a criterion, in this paper we assume pri-
orities in order of vehicles required and total dis-
The vehicle routing problem with time win- tance, respectively.
dows is a logistics and distribution management Both optimization and heuristic approaches
problem that has many real-world applications. have been developed for the VRPTW. Among
Typical applications of the VRPTW include bank the optimization approaches, Desrochers et al.
deliveries, postal deliveries, industrial refuse col- (1992) have extended the maximum problem size
lection, national franchise restaurant deliveries, that can be solved optimally. The problems con-
gas and petroleum deliveries, school bus routing, sist of 100 customers. Fisher et al. (1997) have
and security patrol services. The VRPTW has developed a k-tree relaxation approach that gen-
been extensively studied by academics due to its erated optimal solutions to two of the Solomon
challenging complexity and strategic cost impact. test problems (Solomon, 1987). Kohl and Madsen
The recent emergence of supply chain manage- (1997) used a Lagrangian relaxation of the con-
ment also renews interest in the study of the straint set requiring that each customer must be
VRPTW. serviced. The master problem consists of finding
The VRPTW can be described briefly as fol- the optimal Lagrangian multipliers and the sub-
lows. Let G = (C, A) be a directed graph where problem is a shortest path problem with time
C = {c0, c1, . . ., cn} is a customer vertex set, and windows and capacity constraints. Because of
A = {(ci, cj):i 5 j} is an arc set. Vertex c0 denotes NP-completeness, the majority of research has fo-
a depot at which m identical vehicles are based, cused on the development of heuristics. Solo-
and the remaining vertices of C represent customer monÕs (1987) construction heuristics represent
locations. Each arc (ci, cj) has an associated non- the seminal work on heuristics for the VRPTW.
negative distance dij and a nonnegative travel time Recently, researchers have focused their attention
tij. The VRPTW consists of designing a set of least on the design of efficient metaheuristics. Their ap-
cost vehicle routes such that: proaches include simulated annealing, tabu
search, evolutionary methods, and genetic algo-
(a) Every route starts and ends at the depot. rithms. Rochat and Taillard (1995) used a prob-
(b) Every customer of C (excluding the depot) is abilistic local search which used scatter-search
visited exactly once by exactly one vehicle. type strategies to solve the VRPTW. Potvin
(c) The total demand of any vehicle route does not et al. (1996) employed an efficient tabu search
exceed the vehicle capacity. metaheuristic to find solutions for the VRPTW.
(d) Each customer ci has a required service time Potvin and Bengio (1996) used a genetic search
si. approach to solve the VRPTW. Chiang and
R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622 609

Russell (1997) designed a reactive tabu search to cult to solve. Quality solutions require a meticu-
solve the VRPTW. Taillard et al. (1997) used lous design of the heuristics. We adapt the
tabu search to solve both the vehicle routing strategic framework of the scatter search to the
problem with soft time windows and the VRPTW. Our scatter search implementation pro-
VRPTW; their approach generated several new cedure is summarized in Fig. 1.
best known solutions. Cordeau et al. (2001) pre- The initial reference set is composed of solu-
sented a unified tabu search for both the periodic tions generated using two different construction
and the multi-depot vehicle routing problems heuristics to generate the initial solution. In one
with time windows. Recently, Gehring and Hom- approach, a tabu search route improvement proce-
berger (2002) used parallelization of a two-phase dure (Chiang and Russell, 1997) is embedded with-
evolutionary metaheuristic to achieve robust solu- in the parallel construction process to improve
tions for the VRPTW. solution quality. To increase the diversity of the
solution pool in the reference set, SolomonÕs
sequential insertion heuristic (Solomon, 1987) is
3. Scatter search for the VRPTW also used to generate different solutions. Parallel
route construction differs from sequential route
In addition to the exploration of strategic possi- construction in that all m emerging routes are con-
bilities for the purpose of combining solutions, structed simultaneously rather than one at a time.
scatter search offers a general framework for The reference set consists of a total of b1 best solu-
implementation as follows. tions from the parallel construction heuristic and
b2 best solutions from SolomonÕs sequential inser-
(1) A diversification generator: to generate a col- tion heuristic.
lection of diverse trial solutions using an arbi- To combine solutions from the initial reference
trary trial solution (or seed solution) as an set, common arcs are generated and put in sepa-
input. rate clusters. From there, N partial solutions are
(2) An improvement method: to transform a trial generated from different types of combinations.
solution into one or more enhanced trial Some of the partial solutions might not be feasible.
solutions. For example, a node (a customer) may have sev-
(3) A reference set update method: to build and eral incoming or outgoing arcs. A restoring proce-
maintain a reference set consisting of the b best dure is used to repair the infeasible partial
solutions found. solutions. A parallel construction approach is then
(4) A subset generation method: to operate on the used to connect the remaining nodes and tabu
reference set, to produce a subset of its solu- search is called periodically to improve the par-
tions as a basis for creating combined tially constructed routes.
solutions. The scatter search continues to update the refer-
(5) A solution combination method: to transform ence set. The new reference set is composed of
a given subset of solutions produced by the (b1 + b2)/2 best solutions from the original refer-
subset generation method into one or more ence set, [(b1 + b2)/2]  1 best solutions out of N
combined solution vectors. generated solutions, and 1 new solution from the
set covering approach for the N generated solu-
Laguna (2002), Laguna and Marti (2002a,b), tions. For the updated reference set, the same prin-
Glover et al. (2000a,b), and Glover (1999) provide ciple is used again. Quality and diversity are the
detailed descriptions and an implementation two dominant factors in the reference set design.
framework of the scatter search approach. In more The [(b1 + b2)/2]  1 best solutions out of N gener-
recent work, Marti et al. (2003) provided the main ated solutions added to the reference set should
principles and ideas of scatter search. possess the property of low percentage of arcs
The VRPTW is a highly constrained problem. common to the solutions in the reference to
Time window constraints make the problem diffi- achieve diversity. At the same time, those solutions
610 R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622

Fig. 1. Scatter search procedure for the VRPTW.

should have the least number of routes and dis- that is shared by different solutions. We assume
tance to achieve quality. that arcs that appear with a high frequency in elite
The scatter search procedure stops when a ter- solutions will have a higher probability of being in
mination criterion—the maximum number of iter- an optimal solution. Therefore, we want to create
ations, MaxIter, is reached, the reference set does new solutions whose nucleus consists of arcs com-
not change, or improvement does not warrant fur- mon to subsets of the reference set.
ther iterations. To generate common arcs, we operate on
We discuss below critical components of our our reference solution set RefSet = {s1, s2, s3, . . .,
scatter search procedure. sjRefSetj}. The solutions in RefSet are sorted
according to their objective function values (where
3.1. Combining solutions the best overall solution is first on the list). We
follow GloverÕs approach (Glover, 1999) and gen-
Scatter search operates on a pool of solutions in erate common arc subsets for the following com-
the reference set. New solutions are generated by binations:
applying a solution combination method.
• 2-Solution combination subsets: {s1, s2}, {s1,
3.1.1. Arc combination approach s3}, . . ., {s2, s3}, . . ., {sjRefSetj1, sjRefSetj}.
As we mentioned in the introduction, the solu- • 3-Solution combination subsets: {s1, s2, sb}, {s1,
tion of the VRPTW can be represented as a net- s3, sb}, . . ., {s2, s3, sb}, . . ., {sjRefSetj1, sjRefSetj, sb},
work G = (C, A) where C = {c0, c1, . . ., cn} is a where sb is the best solution (other than the first
customer vertex set, and A = {(ci, cj): i 5 j} is an two solutions). That is, 3-element subsets are
arc set. Each arc (ci, cj) has an associated nonneg- derived from the 2-element subsets by augment-
ative distance dij and a nonnegative travel time ing each 2-element subset to include the best
tij. To combine solutions, we adopt a common solution (as measured by the objective value)
arc strategy. We define a common arc as an arc not in this subset.
R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622 611

• 4-solution combination subsets: {s1, s2, sb1, sb2}, solutions. An elite common arc list is created by
{s1, s3, sb1, sb2}, . . ., {s2, s3,sb1, sb2}, . . ., {sjRefSetj1, adding only arcs which are of sufficient merit.
sjRefSetj, sb1, sb2}, where sb1, sb2 are the two best Given a weight threshold value a, common arcs
solutions (other than the first two). That is, 4- are placed in the elite common arc list if Wrs P a.
element subsets are derived from the 3-element In the course of experimentation, we found that a
subsets by augmenting each 3-element subset threshold values of 1.0, 0.85, 0.85, and 0.80
to include the best solution (as measured by worked well for the 2-solution, 3-solution, 4-solu-
the objective value) not in this subset. tion, and best-i solution combination subsets,
• The subsets consisting of the best i elements (as respectively. The arcs in the common arc list com-
measured by the objective value), for i = 5 to prise a set of separated groups which we call
jRefSetj. ‘‘clusters’’ or seed routes.
To have a feasible solution, the arcs with dupli-
The total number
 of subset
 solutions generated cate heads and tails in the common arc list need to
jRefSetj be modified in order to eliminate subtours. Time
is equal to 3  þ jRefSetj  4 . For a
2 window feasibility is checked and nodes with time
small reference set of size 5, not all of the subsets window violations are truncated. The procedure is
can be unique. Of the 31 subsets generated, only a feasibility restoration heuristic. Given the list of
26 could be unique. For size 10 and larger refer- common arcs grouped in clusters, we try to deter-
ence sets, unique subsets can be achieved. Once mine the feasible chains or subtours among those
solution subsets are generated, we compute arc arcs.
weights for arcs to be considered for use in the The heuristic procedure is listed in Fig. 2.
combined solutions. Consider the arc combination The cluster generation and feasibility restora-
set C(S) and the calculation of arc weight Wrs . De- tion can be illustrated in Fig. 3. The light colored
fine solution Si as a solution in the arc combina- arcs represent arcs that have been eliminated to re-
tion set C(S), i.e., Si 2 C(S). By using distance as store feasibility. The dark nodes and arcs represent
a surrogate objective function, we can calculate a restored feasible arc-node clusters that are the nu-
weight threshold value Wrs. cleus of partially completed routes. The white
nodes represent nodes that are to be added to
P P
S j 2CðSÞ k;l2S j d kl
routes during the solution improvement process.
v0i ¼ P ; solution S i 2 CðSÞ: ð1Þ Parallel construction follows to link unassigned
k;l2S i d kl
nodes and tabu search is used to improve the
v0i emerging routes.
vi ¼ P 0
;
S j 2CðSÞ vj 3.1.2. Route combination approach
vi ¼ relative value of solution S i : ð2Þ In addition to the common arc subset gener-
ation approach, we also examine a global route
X combination approach. Set partitioning has been
W rs ¼ ars vi ;
S i 2CðSÞ
used in a variety of vehicle routing and distribu-
 tion applications. Bramel and Simchi-Levi (1997)
1 if arc ðr; sÞ 2 S i ;
where ars ¼ : ð3Þ examined the effectiveness of set partitioning in
0 otherwise: solving vehicle routing problems with time win-
dows. They observed that for any distribution
Here, Eqs. (1) and (2) calculate the relative value of service times, time windows, customer loads,
of solution i using the total distance as the evalu- and locations, the relative gap between frac-
ation criterion. The lower the total distance solu- tional and integer solutions of the set partition-
tion i has, the higher the relative value it will ing problem becomes arbitrarily small as the
get. Eq. (3) calculates the weight value, or merit, number of customers increases. Ronen (2000)
of a common arc (r, s) as it appears in different used elastic set partitioning to generate solutions
612 R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622

Fig. 2. Restoring feasibility: cluster generation procedure.

X
within 1% of optimality for scheduling charter min zr x r
aircraft. r2R
Set covering is an optimization–based approach s:t:
that is capable of not only handling nonlinear and X
fixed route costs, but is also capable of generating pir xr P 1; i 2 N n fdg; ð4Þ
optimal solutions provided the problem size is not r2R
X
too large. In order to guarantee an optimal solu- xr 6 m; ð5Þ
tion, all feasible routes must be considered implic- r2R
itly or explicitly in the integer programming
formulation. To generate an approximate or heu- xr 2 f0; 1g; r 2 R;
ristic solution, a subset of all possible routes can pir 2 f0; 1g; i 2 N n fdg; r 2 R;
be generated in order to reduce problem size and
computation time. where
The route combination approach uses the (
set covering algorithm to select the minimum 1 customer i is visited by route r;
cost solution from all among all the routes in the pir ¼
0 otherwise;
set of improved trial solutions. To reduce the
size of the resulting integer program, duplicate zr the total cost of route r,
routes are eliminated by sorting and scanning m the specified number of vehicle routes,
the set of all routes in the set of improved R the set of all routes generated by subset
solutions. improvements,
The set covering formulation for the VRPTW is N the depot and customers set,
as follows: d the depot,
R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622 613

Define an arc as a node pair: aij (node i, node j) monÕs sequential insertion heuristic (Solomon,
1987). The same tabu search approach is used
aij a jk
again in the improvement of solutions generated
by the arc combination method. The tabu search
with advanced recovery (Chiang and Russell, in
Customer press) is used to improve the unique solution gen-
erated by the set covering optimization method.
The tabu search metaheuristics employ a lexico-
graphic objective function in which the most
important objective is the number of vehicle routes
Depot required. The secondary objective is a convex com-
bination of total travel distance and time. Both the
reactive tabu search and tabu search with ad-
vanced recovery use the same neighborhood and
tabu data structure which is defined as follows.
The neighborhood structure is defined by the k-
interchange moves attributed to Osman (1993) and
Osman and Salhi (1996). The k-interchange moves
consist of three possible node exchanges: insert,
delete, and swap. The format of the insert move-
ment is insert (r1, r2, i, j). It moves a node from
route r1 position i to route r2 position j. We label
the insert movement as a (0, 1) movement. The for-
mat of the delete movement is delete (r1, r2, i, j). It
moves the node from route r2 position j to route
Fig. 3. Restoring feasibility: an illustration of cluster genera-
r1 position i. The delete movement is also called
tion algorithm.
a (1, 0) movement. The third movement is an ex-
change (swap) movement. The format of the ex-
change movement is exchange (r1, r2, i, j). It

1 route is selected for the final solution; exchanges the node in route r1 position i and route
xr ¼ r2 position j. The exchange movement is labeled as
0 otherwise:
a (1, 1) movement. Not every move is feasible. The
The first constraint set (4), represents ‘‘cover- feasibility for the movements of nodes is dictated
ing’’ type constraints that ensure that each cus- by time windows and the maximum allowable tra-
tomer is served by at least one vehicle. The vel time duration of a route. The neighborhood
second constraint set (5), ensures that the total structure was effective in solving vehicle routing
number of vehicle routes used does not exceed problems with time windows.
the maximum number of vehicles available. The tabu data structure can be illustrated with
the neighborhood structure. A node is said to have
tabu status with respect to a route if it cannot be
3.2. Improvement heuristics switched back to that route. Define TABU(u, r),
"u 2 {1, . . ., jCj}, "r 2 {1, . . . , jRj}, where jCj is
We use two versions of tabu search metaheuris- the cardinality of the set of nodes and jRj is the
tics to improve solutions. A short-term reactive cardinality of the set of routes. TABU(u, r) records
tabu search approach (Chiang and Russell, 1997) the tabu status of node u in route r, after being
is used in the improvement of the initial reference moved to another route, i.e., if the TABU(u, r) = x,
set of solutions generated by the parallel construc- then node u cannot return to route r until iteration
tion method (Chiang and Russell, 1997) and Solo- x finishes. Let t be a tabu list size. If node u
614 R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622

switches from route r to r 0 at the current iteration, quality and the tradeoffs between quality and
then TABU(u, r) is updated as follows: diversity in reference set design.
The test data consists of the well known 56
TABU ðu; rÞ ¼ iter þ t where iter is the
Solomon VRPTW test problems. Each of these
current iteration number; problems has 100 customers and the travel time
Node u is tabu with respect to route r between nodes is equal to the Euclidean distance.
if TABU ðu; rÞ > iter: The 56 test problems are grouped into six problem
types. The geographical distribution of customers
Hence, for k-interchange operators (0, 1) or in problem sets R1 and R2 is randomly generated
(1, 0), if u switches from r1 to r2 and if according to a uniform distribution. Problem sets
TABU(u, r2) > iter, then the move is tabu; other- C1 and C2 have clustered routes based on solu-
wise, it is not tabu. For operator (1, 1), if node u tions to a VRP. Sets RC1 and RC2 are semi-clus-
in route r1 is exchanged with node u 0 in route r2 tered with a mix of randomly distributed and
and if TABU(u, r2) > iter or TABU(u 0 , r1) > iter, clustered customers. Sets R1, C1, and RC1 have
then the move is tabu; otherwise, it is not. When a shorter scheduling horizon and allow fewer cus-
a node is tabu, it cannot return to its original route tomers per route. Problem sets R2, C2, and RC2
in the next Tabu List Size iterations if it switches have a longer scheduling horizon and allow a lar-
from that route to another unless the aspiration ger number of customers per route. A more com-
criterion is satisfied. The aspiration criterion is as plete description of the test problems can be
follows: Tabu status can be overridden if the found in Solomon (1987).
switch of a node in Tabu can result in a new solu- Reference set size is an important factor in both
tion with cost less than the best solution found so the computational requirements and solution qual-
far. ity of a scatter search implementation. Size 20 has
The tabu search with advanced recovery (Chi- been used in previous adaptations of scatter search.
ang and Russell, in press) is applied after a pool In Table 1, we show the solution quality and compu-
of solutions has been generated through tabu tation time required to solve problems in the R1 and
search. Our approach is a variation of the ad- RC1 problem sets.  set of size jRefSetj
 A reference
vanced recovery approach proposed by Xu et al. jRefSetj
(1999). The advanced recovery algorithm records will generate 3  þ jRefSetj 4 subset
2
the 3 best solutions obtained from the tabu search problems to be solved. Thus, reference sets of size
algorithm, then each of the 3 best solutions is used 20, 10, and 5 will generate 586, 141, and 31 sub-
to be the initial solution to start another iteration problems, respectively.
of tabu search. The 9 best solutions generated The results achieved from the three reference set
from the original 3 best solutions are stored in a sizes are similar and this is not surprising given
temporary pool. After the original 3 best solutions that the same tabu search metaheuristic was used
are recovered, the 9 best solutions stored in the to obtain solutions from the subproblems. How-
temporary pool are compared to the original 3 ever, size does facilitate slightly better results.
solutions to get the 3 best ones among the 12. Two fewer vehicles were required using the size
The comparison yields the 3 best solutions for 20 reference set relative to the size 5 reference set
the next iteration. After a given number of itera- and one less than when using the size 10. The total
tions, the algorithm stops. distance was less for size 10, however, some prob-
lems require more distance traveled in order to
achieve fewer vehicles routes. The computational
4. Computational experiments requirements of a reference set of size 20 are signif-
icantly greater, but because of the superior results
In this section we analyze our scatter search and the larger solution pool that is generated, we
implementation for solving the VRPTW. We also chose size 20 as the basis of the scatter search
examine the effects of reference set size on solution implementation.
R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622 615

Table 1
Initial search with different size reference sets
Problem Reference set size 20 Reference set size 10 Reference set size 5
Vehicles Distance Vehicles Distance Vehicles Distance
R101 19 1651.01 19 1652.17 19 1652.17
R102 17 1492.16 17 1487.47 17 1486.65
R103 13 1306.5 13 1307.63 13 1316.38
R104 10 992.3 10 991.1 10 990.21
R105 14 1379.9 14 1384.03 14 1379.26
R106 12 1257.88 12 1259.91 13 1239.98
R107 10 1128.45 10 1122.67 10 1115.53
R108 9 1005.3 9 978.55 9 976.24
R109 12 1165.71 12 1167.24 12 1158.34
R110 11 1092.36 11 1095.84 11 1104.29
R111 10 1106.45 10 1120.26 10 1118.48
R112 10 975.04 10 971.35 10 969.74
RC101 14 1769.34 15 1633.3 15 1641.33
RC102 13 1485.32 13 1485.7 13 1524.64
RC103 11 1266.3 11 1281.53 11 1293.54
RC104 10 1138.02 10 1145.6 10 1167.97
RC105 13 1650.12 13 1648.43 14 1549.46
RC106 12 1392.6 12 1396.2 12 1414.49
RC107 11 1237.31 11 1245.7 11 1257.32
RC108 10 1149.56 10 1149.56 10 1181.43
Total 241 25642 242 25524 243 25654
CPU timea 3063 778 163
a
Minutes Athlon 2.25 GHz processor.

Updating the reference set requires that special search, we devised three strategies for reducing
attention be given to solution diversity as well as the computational requirements of the scatter
quality. Simply selecting the best 20 solutions from search implementation. First, the tabu search
the union of the existing reference set and the 586 improvement metaheuristic was modified to have
generated solutions yielded a reference set that was only 10 iterations per customer or a maximum of
comprised of too many similar solutions. We 500 iterations. Second, the set of 586 subproblems
incorporated a design strategy based on the per- is comprised of a significant number of duplicate
centage of arcs in a solution that were already in subproblems. We implemented a reduction policy
the existing reference set. The update strategy that allows only one subproblem for a given num-
was based on selecting new solutions that had a ber of initial clusters and associated number of
lower percentage of common arcs but that were customers. This policy eliminates some non-redun-
constrained to be within 2–3% of the best solution dant subproblems but is necessary to facilitate
objective value in the reference set. The results in computational tractability. Thirdly, we observed
Table 1 used the full set of 586, 141, and 31 subsets that subproblems having too few or too many
of subproblems, respectively. The tabu search clusters generally led to inferior solutions. If v
improvement heuristic was long term in that it equals the target number of vehicle routes, then
used a maximum of 1000 iterations. the subproblems having initial routes outside the
range of [v  1, v + 5] were eliminated. Far fewer
4.1. Subset reduction than 586 subproblems were improved in the fol-
lowing experiments reported in Tables 2–6. The
Given both the large subproblem set size of 586 number of subproblems ranged from 35 to 274
and the computational time requirements of tabu out of a possible 586. Subset reduction has little
616 R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622

Table 2
Scatter search for R1, RC1, R2, and RC2: iteration 1
Problem Tabu improvement Set covering Advance recovery
NR Distance CPU time NR Distance CPU time NR Distance CPU time
R101 19 1650.80 11.81 19 1650.80 0.00 19 1650.80 0.18
R102 17 1486.12 15.69 17 1486.12 0.00 17 1486.12 0.58
R103 13 1297.05 22.85 13 1294.67 0.02 13 1294.67 0.22
R104 10 993.78 48.84 10 984.69 0.19 10 984.69 0.18
R105 14 1377.80 20.07 14 1377.11 0.08 14 1377.11 0.16
R106 12 1254.27 19.73 12 1253.05 0.05 12 1253.05 0.18
R107 10 1121.30 19.97 10 1112.42 0.02 10 1106.15 0.22
R108 9 969.67 44.67 9 969.67 1.33 9 969.67 0.20
R109 11 1215.12 25.72 11 1194.73 0.24 11 1194.73 0.15
R110 11 1103.38 32.33 10 1145.68 6.28 10 1145.68 0.19
R111 10 1123.43 25.60 10 1103.68 0.07 10 1103.68 0.17
R112 10 979.96 27.56 10 968.21 1.75 10 968.21 0.15
RC101 14 1743.96 23.85 14 1695.94 0.10 14 1695.94 0.15
RC102 12 1595.34 26.75 12 1558.15 1.21 12 1558.15 0.18
RC103 11 1269.38 31.97 11 1263.63 0.12 11 1263.63 0.21
RC104 10 1140.08 39.08 10 1137.54 0.32 10 1137.54 0.13
RC105 13 1641.06 27.30 13 1629.44 0.03 13 1629.44 0.13
RC106 12 1394.50 22.71 11 1494.14 1.11 11 1494.14 0.17
RC107 11 1234.49 31.43 11 1230.54 0.11 11 1230.54 0.12
RC108 10 1160.84 40.34 10 1152.76 0.45 10 1152.76 0.12
R201 4 1279.37 10.27 4 1274.22 0.06 4 1258.64 0.18
R202 3 1262.74 17.88 3 1262.74 0.02 3 1240.47 0.21
R203 3 959.71 11.86 3 959.71 0.01 3 955.87 0.30
R204 2 857.52 23.41 2 857.52 0.01 2 856.35 0.36
R205 3 1087.37 2.79 3 1087.37 0.00 3 1043.90 0.20
R206 3 949.02 21.74 3 949.02 0.01 3 947.32 0.32
R207 2 928.15 21.68 2 928.15 0.01 2 928.15 0.21
R208 2 753.37 15.87 2 753.37 0.00 2 742.32 0.63
R209 3 951.41 11.18 3 951.41 0.00 3 945.71 0.30
R210 3 1005.90 6.82 3 1005.9 0.00 3 998.69 0.31
R211 3 836.42 14.04 3 836.42 0.03 3 799.20 0.44
RC201 4 1473.95 2.99 4 1473.95 0.01 4 1447.70 0.17
RC202 3 1492.59 5.07 3 1492.59 0.00 3 1467.43 0.18
RC203 3 1131.25 6.65 3 1131.25 0.00 3 1082.37 0.29
RC204 3 824.51 17.77 3 824.51 0.00 3 820.89 0.41
RC205 4 1355.70 4.49 4 1355.70 0.00 4 1355.70 0.16
RC206 3 1179.37 8.75 3 1179.37 0.00 3 1178.27 0.21
RC207 3 1232.21 9.83 3 1232.21 0.00 3 1164.34 0.28
RC208 3 887.89 28.01 3 887.89 0.01 3 881.34 0.43

effect on solution quality while significantly reduc- search columns show the number of vehicle routes
ing computation time. required and total distance of the best solution
Tables 2 and 3 show the results of the first iter- found among all subproblems generated. The
ation of the scatter search implementation. Table 2 CPU time is the total for all subproblems im-
reports the results pertaining to the R1, RC1, R2, proved. The set covering columns show the result
and RC2 problem sets while Table 3 shows the re- from combining all solutions into one ‘‘elite’’ solu-
sults for problem sets C1 and C2. The details show tion. Recent advances in integer programming
the solution properties of the initial search and two solvers such as CPLEX 8.1 enable relatively large
subsequent improvement methods. The initial 0–1 models with up to 16,000 variables and 101
R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622 617

Table 3
Scatter search for C1, and C2: iteration 1
Problem Tabu improvement Set covering Advance recovery
NR Distance CPU time NR Distance CPU time NR Distance CPU time
C101 10 828.94 13.78 10 828.94 0.00 10 828.94 0.10
C102 10 828.94 8.18 10 828.94 0.15 10 828.94 0.12
C103 10 824.78 6.32 10 824.78 0.64 10 824.78 0.16
C104 10 828.94 61.38 10 828.94 0.98 10 828.94 0.23
C105 10 828.94 17.70 10 828.94 0.00 10 828.94 0.11
C106 10 828.94 27.01 10 828.94 0.00 10 828.94 0.11
C107 10 828.94 18.13 10 828.94 0.00 10 828.94 0.13
C108 10 828.94 6.94 10 828.94 0.10 10 828.94 0.14
C109 10 828.94 4.91 10 828.94 0.31 10 828.94 0.21
C201 3 591.56 17.56 10 591.56 0.00 10 591.56 0.11
C202 3 591.56 28.16 3 591.56 0.05 3 591.56 0.19
C203 3 591.17 64.91 3 591.17 0.05 3 591.17 0.31
C204 3 599.68 13.74 3 590.60 0.75 3 590.60 0.23
C205 3 588.88 30.56 3 588.88 0.00 3 588.88 0.14
C206 3 588.49 4.49 3 588.49 0.00 3 588.49 0.18
C207 3 588.49 6.62 3 588.49 0.43 3 588.49 0.18
C208 3 588.32 25.28 3 588.32 0.54 3 588.32 0.27

constraints to be solved in minutes if not seconds. Table 6 shows the solution quality generated by
In 2 of the 56 test problems, set covering was able our scatter search implementation, Rochat and
to reduce the number of vehicles required relative TaillardÕs approach, and the best known results
to the initial search. (optimal and heuristic) from many researchers
The advanced recovery columns show the re- as reported on SolomonÕs website, http://web.
sults of applying tabu search with advanced recov- cba.neu.edu/~msolomon/problems.htm. The total
ery to the single solution obtained by set covering. accumulated CPU time for the scatter search solu-
The computing times are small because the start- tion process is shown in Table 6.
ing solution is an elite solution and the maximum Table 6 also shows the gap between the best
iterations of the tabu search is limited to 1000. Ad- known solution and our scatter search. The gap
vanced recovery did not eliminate any vehicle is reported as the percentage increase in the
routes but did achieve distance reduction in some objective function relative to the best known.
instances. For problems such as R205, the For the 4 problems in which scatter search did
improvement was significant. not find the minimum number of vehicles, the
Tables 4 and 5 show the results obtained in the gap is not reported. A zero gap indicates the
second iteration of scatter search. The second attainment of a best known solution. The average
search columns show the details of the solution ob- gap is only 1.060% with a maximum gap of
tained from the updated reference set. In Table 4, 5.93%. These percentages indicate that scatter
there were 22 instances out of 39 in which the solu- search provides a framework for a robust solu-
tion improved relative to the initial search. In Table tion methodology for the VRPTW. In Table 6
5 there were no improvements, however, in the first we include the results obtained by Rochat and
iteration, the scatter search had already found the Taillard (1995) since their work most closely
best known heuristic solutions. The solution pool resembles a scatter search implementation for
used for set covering includes the subproblem solu- the VRPTW. In their probabilistic local search,
tions from both iterations plus the best solution they employ a pool of solutions similar to a
found in iteration 1. At the completion of iteration reference set and apply local search to improve
2, several best known solutions had been achieved. solutions. They also ‘‘combine’’ solutions by
618 R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622

Table 4
Scatter search for R1, RC1, R2, and RC2: iteration 2
Problem Tabu improvement Set covering Advance recovery
NR Distance CPU time NR Distance CPU time NR Distance CPU time
R101 19 1650.80 1.13 19 1650.80 0.01 19 1650.8 –a
R102 17 1486.12 1.86 17 1486.12 0.01 17 1486.12 –a
R103 13 1298.34 3.76 13 1293.59 0.04 13 1292.94 0.39
R104 10 989.40 15.94 10 982.02 0.18 10 982.02 0.19
R105 14 1377.11 12.03 14 1377.11 0.03 14 1377.11 –a
R106 12 1254.27 2.14 12 1252.06 0.13 12 1252.06 0.18
R107 10 1121.70 10.14 10 1104.96 1.13 10 1104.96 0.19
R108 9 968.63 14.20 9 968.63 4.80 9 968.63 0.19
R109 11 1195.36 8.37 11 1194.73 0.67 11 1194.73 0.20
R110 11 1094.91 17.00 10 1122.55 12.52 10 1121.01 0.31
R111 10 1115.58 15.50 10 1102.82 0.90 10 1102.82 0.19
R112 10 967.59 20.06 10 960.78 5.31 10 960.78 0.18
RC101 14 1698.82 11.30 14 1696.95 0.04 14 1696.95 0.15
RC102 12 1602.16 10.11 12 1557.56 13.00 12 1557.56 0.16
RC103 11 1264.59 9.71 11 1261.69 0.86 11 1261.69 0.14
RC104 10 1138.59 13.14 10 1135.53 1.43 10 1135.53 0.17
RC105 13 1639.54 1.42 13 1629.44 0.08 13 1629.44 0.13
RC106 12 1392.11 13.43 11 1428.77 0.39 11 1425.77 0.25
RC107 11 1232.87 10.17 11 1230.54 0.34 11 1230.54 0.12
RC108 10 1146.66 13.54 10 1146.66 1.66 10 1146.66 0.13
R201 4 1277.64 12.24 4 1258.64 0.04 4 1258.64 –a
R202 3 1228.54 16.50 3 1228.54 0.04 3 1228.54 0.21
R203 3 961.72 17.79 3 955.87 0.03 3 955.31 0.67
R204 2 867.13 26.36 2 856.35 0.03 2 850.70 0.98
R205 3 1079.05 19.85 3 1043.90 0.02 3 1021.88 1.00
R206 3 949.02 26.20 3 947.32 0.14 3 912.26 0.86
R207 2 928.15 16.21 2 928.15 0.03 2 928.15 –a
R208 2 755.15 45.47 2 742.32 0.02 2 741.33 0.79
R209 3 962.88 13.56 3 945.71 0.13 3 945.71 –a
R210 3 993.71 15.55 3 993.71 0.02 3 983.66 0.30
R211 3 808.07 51.05 3 799.20 0.02 3 796.41 0.92
RC201 4 1479.63 2.76 4 1435.17 0.06 4 1435.17 0.90
RC202 3 1564.15 3.34 3 1467.43 0.01 3 1448.15 0.13
RC203 3 1139.21 8.97 3 1082.37 0.00 3 1082.37 0.12
RC204 3 833.89 54.85 3 820.88 0.01 3 820.88 0.41
RC205 4 1351.6 15.32 4 1338.29 0.05 4 1338.28 0.20
RC206 3 1198.55 8.28 3 1178.27 0.06 3 1178.27 0.13
RC207 3 1107.27 18.71 3 1107.26 0.05 3 1076.62 1.20
RC208 3 859.44 48.80 3 859.44 0.05 3 834.27 0.99

constructing new ones from components of previ- 43.33, 163.33, 130, 53.33, and 120 minutes for
ous solutions. Our scatter search implementation problem sets R1, RC1, R2, RC2, C1, and C2,
was able to achieve a reduction of 8 vehicles rel- respectively, using a 64 bit Silicon Graphics
ative to Rochat and Taillard and with a slight in- Indigo 100 Mhz computer. On the same problem
crease in total distance. Comparing computation sets our scatter search implementation required
times is difficult since the Rochat and Taillard re- an average of 39.78, 43.75, 38.93, 31.39, 17.74,
sults are the best found in multiple experiments. and 22.14 minutes, respectively, on a 32 bit Ath-
However, they report average CPU times of 45, lon 2.25 Ghz computer.
R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622 619

Table 5
Scatter search for C1, and C2: iteration 2
Problem Tabu improvement Set covering Advance recovery
NR Distance CPU time NR Distance CPU time NR Distance CPU time
C101 10 828.94 12.01 10 828.94 0.02 10 828.94 –
C102 10 828.94 6.95 10 828.94 0.18 10 828.94 –
C103 10 824.78 7.67 10 824.78 0.66 10 824.78 –
C104 10 828.94 49.87 10 828.94 0.99 10 828.94 –
C105 10 828.94 15.29 10 828.94 0.01 10 828.94 –
C106 10 828.94 29.53 10 828.94 0.01 10 828.94 –
C107 10 828.94 18.69 10 828.94 0.02 10 828.94 –
C108 10 828.94 5.43 10 828.94 0.14 10 828.94 –
C109 10 828.94 3.84 10 828.94 0.33 10 828.94 –
C201 3 591.56 15.16 10 591.56 0.03 10 591.56 –
C202 3 591.56 22.28 3 591.56 0.06 3 591.56 –
C203 3 591.17 50.71 3 591.17 0.07 3 591.17 –
C204 3 599.68 10.23 3 590.60 0.78 3 590.60 –
C205 3 588.88 25.52 3 588.88 0.01 3 588.88 –
C206 3 588.49 3.93 3 588.49 0.01 3 588.49 –
C207 3 588.49 7.88 3 588.49 0.43 3 588.49 –
C208 3 588.32 21.94 3 588.32 0.55 3 588.32 –
–: indicates solution found in previous iteration and no CPU time required.

Table 6
Scatter search solution quality
Problem Best known Rochat and Tallard Scatter search solutions % Gap
NR Distance NR Distance NR Distance CPU time
R101 18 1607.7 19 1650.8 19 1650.82 13.13 –
R102 17 1434 17 1486.12 17 1486.12 18.14 3.635
R103 13 1292.68 14 1217.44 13 1292.94 27.28 0.020
R104 9 1007.24 10 984.69 10 982.18 65.52 –
R105 14 1377.11 14 1377.11 14 1377.11 32.37 0
R106 12 1251.98 12 1251.98 12 1252.06 22.41 0.006
R107 10 1104.66 10 1159.86 10 1104.66 31.67 0
R108 9 960.88 9 997.33 9 968.63 65.39 0.807
R109 11 1194.73 11 1245.65 11 1194.73 35.35 0
R110 10 1118.59 11 1082.05 10 1121.01 68.63 0.216
R111 10 1096.72 10 1184.22 10 1102.82 42.43 0.556
R112 9 982.14 10 953.56 10 960.78 55.01 –
RC101 14 1696.94 15 1623.61 14 1696.94 35.59 0
RC102 12 1554.75 13 1477.56 12 1557.56 51.41 0.181
RC103 11 1261.67 11 1262.01 11 1261.69 43.01 0.001
RC104 10 1135.48 10 1135.84 10 1135.53 54.27 0.004
RC105 13 1629.44 14 1540.24 13 1629.44 29.09 0
RC106 11 1424.73 12 1384.95 11 1425.77 38.06 0.073
RC107 11 1230.48 11 1230.98 11 1230.54 42.29 0.004
RC108 10 1139.82 10 1170.72 10 1146.66 56.24 0.600
R201 4 1252.37 4 1281.55 4 1258.64 22.79 0.501
R202 3 1191.7 4 1088.07 3 1228.54 34.86 3.091
R203 3 939.54 3 948.7 3 955.31 30.66 1.678
R204 2 825.52 2 869.24 2 850.7 51.15 3.050
R205 3 994.42 3 1063.24 3 1021.88 23.86 2.761
(continued on next page)
620 R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622

Table 6 (continued)
Problem Best known Rochat and Tallard Scatter search solutions % Gap
NR Distance NR Distance NR Distance CPU time
R206 3 906.14 3 912.94 3 912.26 49.27 0.675
R207 2 893.33 3 814.78 2 928.15 38.15 3.898
R208 2 726.75 2 738.55 2 741.33 62.78 2.006
R209 3 909.16 3 944.65 3 945.71 25.17 4.020
R210 3 939.34 3 967.45 3 983.66 23 4.718
R211 2 892.71 2 949.47 3 796.41 66.5 –
RC201 4 1406.91 4 1438.94 4 1435.17 6.88 2.009
RC202 3 1367.09 4 1165.56 3 1448.15 8.73 5.929
RC203 3 1049.62 3 1079.58 3 1082.37 16.03 3.120
RC204 3 798.41 3 806.75 3 820.88 73.45 2.814
RC205 4 1297.19 4 1333.71 4 1338.28 20.22 3.168
RC206 3 1146.32 3 1212.68 3 1178.27 17.43 2.787
RC207 3 1061.14 3 1085.63 3 1076.62 30.08 1.459
RC208 3 828.14 3 833.95 3 834.27 78.29 0.740
C101 10 827.3 10 829.01 10 828.94 25.91 0.198
C102 10 827.3 10 829.01 10 828.94 15.58 0.198
C103 10 826.3 10 828.14 10 828.06 15.45 0.212
C104 10 822.9 10 841.66 10 824.78 113.45 0.228
C105 10 827.3 10 829.01 10 828.94 33.11 0.198
C106 10 827.3 10 829.01 10 828.94 56.66 0.198
C107 10 827.3 10 829.01 10 828.94 36.97 0.198
C108 10 827.3 10 829.01 10 828.94 12.75 0.198
C109 10 827.3 10 829.01 10 828.94 9.6 0.198
C201 3 589.1 3 591.58 3 591.56 32.86 0.417
C202 3 589.1 3 591.58 3 591.56 50.74 0.417
C203 3 588.7 3 591.20 3 591.17 116.05 0.419
C204 3 590.6 2 597.78 3 590.60 25.73 0
C205 3 586.4 3 588.90 3 588.88 56.23 0.423
C206 3 586.0 3 588.52 3 588.49 8.61 0.425
C207 3 585.8 3 588.52 3 588.49 15.54 0.459
C208 3 585.8 3 588.52 3 588.32 48.58 0.430
Total 404 57070 416 57151 408 57589 2180.40
Average gap 1.060

Source: Marius M. Solomon http://web.cba.neu.edu/~msolomon/problems.htm.

5. Conclusions Conducting the experiments has enabled us to


make several observations. In solving the highly
We have utilized a scatter search framework to constrained VRPTW, fast local search procedures
solve the VRPTW. The scatter search framework employed in the scatter search framework failed to
provided a means to combine solutions, diversify, yield solution quality competitive with the best
and intensify the metaheuristic search process. metaheuristics reported in the literature. However,
Both the common arc method and set covering utilizing a short-term reactive tabu search meta-
are useful ways to combine vehicle routing solu- heuristic together with a set covering optimiza-
tions. The proposed scatter search method was tion-based procedure has led to a robust though
successful in that 6 best known solutions were ob- somewhat computationally expensive solution
tained in solving the 56 test problems. Addition- method. The proposed method is more efficient
ally, the objective function gap from the best when applied to routing problems that do not have
known solutions averaged only 1.060%. a large number of customers per route. Computa-
R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622 621

tional requirements are significantly reduced by windows. The Journal of the Operational Research Society
subset reduction methods that eliminate duplicate 52 (8), 928–936.
Cung, V., Mautor, T., Michelon, P., Tavares, A., 1997. A
subproblems as well as subproblems unlikely to scatter search based approach for the quadratic assignment
yield an effective solution. problem. In: Proceedings of the IEEE-ICECÕ97 Conference
We found that the size of the reference set does in Indianapolis, April 13–16.
affect solution quality. Using 20 solutions in the Desrochers, M., Desrosiers, J., Solomon, M., 1992. A new
reference set achieved fewer vehicle routes com- optimization algorithm for the vehicle routing problem with
time windows. Operations Research 40, 342–354.
pared to size 5 or size 10. Diversity must be a con- Fisher, M., Jörnsten, K.O., Madsen, O.B.G., 1997. Vehicle
sideration in reference set design. We found that routing with time windows: Two optimization algorithms.
simply selecting the best 20 solutions found to up- Operations Research 45 (3), 488–492.
date the reference set (at least in this application) Gehring, H., Homberger, J., 2002. Parallelization of a two-
yielded mediocre results with the search process phase metaheuristic for routing problems with time win-
dows. Journal of Heuristics 8 (3), 251–277.
limited to a smaller solution space. Glover, F., 1998. A template for scatter search and path
We have shown that a scatter search framework relinking. In: Hao, J.-K., Lutton, E., Ronald, E., Schoen-
can be used to generate solution quality that is auer, M., Snyers, D. (Eds.), Artificial Evolution, Lecture
competitive with the current best context depend- Notes in Computer Science, vol. 1363. Springer, pp. 13–54.
ent metaheuristics. The challenge for the future is Glover, F., 1999. Scatter search and path relinking. In: Corne,
M., Dorigo, M., Glover, F. (Eds.), New Ideas in Optimi-
to develop solution improvement techniques that zation. McGraw Hill, pp. 297–316.
when employed within the scatter search frame- Glover, F., Laguna, M., Marti, R., 2000a. Fundamentals of
work, are as robust and even more efficient. scatter search and path relinking. Control and Cybernetics
39 (3), 653–684.
Glover, F., Laguna, M., Marti, R., 2000b. Scatter search. In:
Ghosh, A., Tsutsui, S. (Eds.), Theory and Applications of
Acknowledgement Evolutionary Computation: Recent Trends. Springer-
Verlag.
The authors are grateful to our graduate assist- Glover, F., Laguna, M., Marti, R., 2002. Scatter search and
ant, Xuan Shi, for his assistance on this project. path relinking: Advances and applications. Working
Paper, The University of Colorado and Universitat de
Valencia.
Glover, F., Løkketangen, A., Woodruff, D., 1999. Scatter
References search to generate diverse MIP solutions. In: Laguna, M.,
Gonazalez-Velarde, J.L. (Eds.), OR Computing Tools for
Bramel, J., Simchi-Levi, D., 1997. The Logic of Logistics. Modeling, Optimization and Simulation: Interfaces in
Springer-Verlag, New York. Computer Science and Operations Research. Kluwer Aca-
Chiang, W.C., Russell, R.A., 1997. A reactive tabu search demic Publishers, pp. 299–317.
metaheuristic for the vehicle routing problem with time Greistorfer, P., 2003. A tabu scatter search metaheuristic for the
windows. INFORMS Journal on Computing 9, 417–430. arc routing problem. Computers & Industrial Engineering
Chiang, W.C., Russell, R.A., A metaheuristic for the vehicle 44 (2), 249–266.
routing problem with soft time windows. The Journal of the Hamiez, J.P., Hao, J.K., 2002. Scatter search for graph
Operational Research Society, in press. coloring. Lecture Notes in Computer Science, vol. 2310.
Campos, V., Laguna, M., Marti, R., 1999. Scatter search for Springer, pp. 168–179.
the linear ordering problem. In: Corne, D., Dorigo, M., Jain, A.S., Meeran, S., 2002. A multi-level hybrid framework
Glover, F. (Eds.), New Ideas in Optimization. McGraw applied to the general flow-shop scheduling problem.
Hill, pp. 331–340. Computers & Operations Research 29 (13), 1873–1901.
Campos, V., Laguna, M., Marti, R., Context-independent Kohl, N., Madsen, O.B.G., 1997. An optimization algorithm
scatter search and tabu search for permutation problems. for the vehicle routing problem with time windows based on
INFORMS Journal on Computing, in press. lagrangian relaxation. Operations Research 45 (3), 395–406.
Corberan, A., Fernandez, E., Laguna, M., Marti, M., 2002. Laguna, M., 2002. Scatter search. In: Pardalos, P.M., Resende,
Heuristic solutions to the problem of routing school buses M.G.C. (Eds.), Handbook of Applied Optimization. Oxford
with multiple objectives. The Journal of the Operational University Press, pp. 183–193.
Research Society 53 (4), 427–435. Laguna, M., Marti, R., 2002a. GRASP and path relinking for
Cordeau, J.-F., Laporte, G., Mercier, A., 2001. A unified tabu 2-layer straight line crossing minimization. INFORMS
search heuristic for vehicle routing problems with time Journal on Computing 11 (1), 44–52.
622 R.A. Russell, W.-C. Chiang / European Journal of Operational Research 169 (2006) 606–622

Laguna, M., Marti, R., 2002b. Scatter Search: Methodology Rego, C., 1998. A Subpath Ejection Method for the Vehicle
and Implementations in C. Kluwer Academic Publishers, Routing Problem. Management Science 44 (10), 1447–1459.
Boston. Rego, C., Spring 2000. Scatter search for vehicle routing
Marti, R., Laguna, M., Glover, F., 2003. Principles of scatter problems. National INFORMS Meeting, Salt Lake City,
search. Working Paper, Departamento de Estadı́stica e UT.
Investigación Operativa, Universitat de Valencia, Spain. Rochat, Y., Taillard, E.D., 1995. Probabilistic diversification
Osman, I.H., 1993. Metastrategy simulated annealing and tabu and intensification in local search for vehicle routing.
search algorithms for the vehicle routing problem. Annals of Journal of Heuristics 1, 147–167.
Operations Research 41, 421–451. Ronen, D., 2000. Scheduling charter aircraft. Journal of
Osman, I.H., Salhi, S., 1996. Local search strategies for the Operational Research Society 51, 258–262.
vehicle fleet mix problem. In: Rayward-Smith, V.J., Osman, Solomon, M., 1987. Algorithms for the vehicle routing and
C.R., Reeves, C.R., Smith, G.D. (Eds.), Modern Heuristic scheduling problem with time window constraints. Opera-
Search Methods. John Wiley & Sons Ltd, pp. 131–154. tions Research 35, 254–265.
Potvin, J.Y., Bengio, S., 1996. The vehicle routing problem with Taillard, E.D., Badeau, P., Gendreau, M., Guertin, F., Potvin,
time windows—Part II: Genetic search. INFORMS Journal J.Y., 1997. A tabu search heuristic for the vehicle routing
on Computing 8, 165–172. problem with soft time windows. Transportation Science 31,
Potvin, J.Y., Kervahut, T., Garcia, B.L., Rousseau, J.M., 1996. 170–186.
The vehicle routing problem with time windows—Part I: Xu, J., Chin, S.Y., Glover, F., 1999. Optimizing a ring-based
Tabu search. INFORMS Journal on Computing 8, 158– private line telecommunication network using tabu search.
164. Management Science 45, 330–345.

You might also like