A Guide to Vehicle Routing Heuristics

Author(s): J-F Cordeau, M. Gendreau, G. Laporte, J-Y Potvin, F. Semet

Reviewed work(s):
Source: The Journal of the Operational Research Society, Vol. 53, No. 5 (May, 2002), pp. 512522
Published by: Palgrave Macmillan Journals on behalf of the Operational Research Society
Stable URL: http://www.jstor.org/stable/823019 .
Accessed: 15/12/2011 05:13
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Palgrave Macmillan Journals and Operational Research Society are collaborating with JSTOR to digitize,
preserve and extend access to The Journal of the Operational Research Society.

http://www.jstor.org

Journal of the Operational

Research

Society

(2002) 53, 512-522

oe,2002 Operational Research Society Ltd. All rights reserved. 0160-5682/02 $15.00

www.palgrave-journals.com/jors

A guide to vehicle routing heuristics

J-F Cordeau,1M Gendreau,2G Laporte,1*J-Y Potvin2 and F Semet3
'Ecole des Hautes Etudes Commerciales,Montreal, Canada, 2Universite de Montreal, Montreal, Canada; and 3Universitede
Valencienneset du Hainaut Cambresis, Valenciennes,France

and
classicaland modemheuristicsfor the vehicleroutingproblemare summarized
Severalof the most important
resultsarereported.
usingfourcriteria:accuracy,speed,simplicityandflexibility.Computational
compared

Journal of the OperationalResearchSociety (2002) 53, 512-522. DOI: 10.1057/palgrave/jors/2601319

Keywords:vehicleroutingproblem;heuristics

Introduction
The Vehicle Routing Problem (VRP), introduced by Dantzig

and Ramserl in 1959, holds a centralplace in distribution

managementand has become one of the most widely studied
problemsin combinatorialoptimization.The Classical VRP
can be formally defined as follows. Let G (V,A) be a
graph where V = {v0vl, ..., v} is a vertex set, and
A = {(vi, j) : vi, Vj E V, i j} is an arc set. Vertex v0 represents a depot, while the remainingvertices correspondto
customers.With A are associated a cost matrix (cij) and a
travel time matrix (tij). If these matrices are symmetrical,
as is commonly the case, then it is standardto define the
VRP on an undirected graph G = (V, E), where
E = {(vi, vj) : vi, vj E V, i < j

is an edge set. Each custo-

mer has a non-negativedemandqi and a service time ti. A

fleet of m identical vehicles of capacity Q is based at the
depot. The numberof vehicles is eitherknown in advanceor
treatedas a decision variable.The VRP consists of designing a set of at most m deliveryor collection routes such that
(1) each routestartsand ends at the depot, (2) each customer
is visited exactly once by exactly one vehicle, (3) the total
demand of each route does not exceed Q, (4) the total
durationof each route (including travel and service times)
does not exceed a preset limit D, and (5) the total routing
cost is minimized. A common variant is where a time
window [a,, bi] is imposed on the visit of each customer.
Severalother extensionshave also been studied;the vehicle
fleet may be heterogeneous,2vehicles may perform both
collections and deliverieson the same route,3some vehicles
may be unable to visit certain sites,4 some customersmay
requireseveral visits over a given time period,5there may
exist more than one depot,5 deliveries may be split among
*Correspondence: G Laporte, Centre de recherche sur les transports
(C.R.T.), Campus de I'Universite de Montreal, C.P 6128, Succursale
Centre-ville, Montreal H3C 3J7, Canada.

severalvehicles,6 etc. For overview articleson the VRP, see

Golden and Assad,7 Fisher,8Desrosiers et al,9 Crainic and
Laporte,'oTothand Vigo,1' Laporteand Semet,12Gendreau
et al13 and Cordeau et al.14

The VRP is a hard combinatorialoptimizationproblem

and only relativelysmall instances can be solved to optimality.To this day, it seems thatno exact algorithmis capable
of consistently solving instances in excess of 50 customers.15This is due to the fact that sharp lower bounds on
the objective value are hard to derive, which means that
partial enumerationbased exact algorithms(using branchand-bound or dynamic programming)will have a slow
convergence rate. Since exact approaches are in general
inadequate,heuristicsare commonly used in practice.
Therehas been a steady evolution,over the past 40 years,
in the developmentof VRP heuristics.In the early classical
heuristics, much of the emphasis was put on quickly
obtaining a feasible solution and possibly applying to it a
postoptimizationprocedure.This class of methods includes
the well-known savings algorithm,16the sweep algorithm17
and the Fisher and Jaikumaralgorithm.18Over the last ten
years, much of the researcheffort has concentratedon the
developmentof algorithmsbased on metaheuristics,using
mainly two principles: local search and population search.

In local search methods, an intensive exploration of the

solution space is performedby moving at each step fromthe
currentsolution to anotherpromising solution in its neighbourhood. Simulated annealing (SA)19 and tabu search

(TS)20are two prime examples of this principle.Population

search consists of maintaininga pool of good parentsolutions and recombiningthem to produceoffspring.A classical example is genetic search (GS)21which combines two
parents to produce offspring. Adaptive memory procedures

(AMPs)22 can be viewed as an extension of GS where

several parents are used to produce several offspring.
While more time consuming than the early heuristics,

JFCordeau
etal-Guide
tovehicle
heuristics
513
routing

metaheuristicsare capable of consistently producing high

quality solutions. Thus, on the Christofides,Mingozzi and
Toth (CMT)23 14 classical VRP benchmarkinstances, the
most powerfulmethodshave producedone provenoptimum
while the solutions obtainedon the remaininginstancesare
believed to be optimal or near-optimalsince they have not
been improvedafterthousandsof hoursof computingeffort,
using sophisticatedsearch techniques.
Yet most commercial software and several in-house
computerprogramsused by companies are based on unsophisticated methodologies, sometimes dating back to the
1960s. Thereare severalreasonsfor this stateof affairs.One
is that the optimizationcomponentof VRP softwareis only
a small part of the product, most of the effort being
expandedon data managementand sophisticateduser interfaces. It is not uncommon for routing experts to use low
qualitysolutions producedby VRP softwareand exploit the
interactive capabilities of the system to perform manual
improvements.Anotherreasonis thatcompanyanalystsand
software developers are simply unawareof the latest algorithmicdevelopments.While there is some truthto this, we
believe the main hindranceto technology transfermay be
more deep-rooted.In this paper,we argue that most of the
available VRP heuristics lack some of the necessary attributes to ensure their adoptionby practitioners.
In the following section we describewhat we believe are
four essential attributesfor softwaretransferabilityand enduser adoption.We then provide an appraisalof some of the
best known heuristicswith respectto these criteria.As much
as we have tried to base this study on published and
objective data, our analysis is at times personaland critical,
especially with respect to the more intangiblecriteria.The
algorithmsaregroupedin two categories:classical heuristics
and metaheuristics.A summary and conclusion close the
paper.

et al,25 can lead to vastly differentresults. Tests are often

carried out by rounding costs d digits after the decimal
point, ie, by setting c, := 10-" LlOdci + 0.5], where L[y is
the integerpartof v. Morerarely,costs aretruncatedd digits
afterthe decimalpoint, ie c := l10- 10dClj. Testscan also
be performedwith floating point arithmetic,without rounding or truncating,using as many digits as allowed by the
computer. When rounding or truncating occurs, it is
common to recompute the final solution cost using more
than d digits afterthe decimal point and to presentthe final
resultwith one or two digits afterthe point. To illustratethe
discrepancyin results,considerthe firstinstanceof the CMT
test problems.The same tabu search algorithm(Taburoute)
yields a cost of 524.61 if computationsare performedwith
roundedcosts c,i and d = 5 (this is, by the way, a proven
optimum26),a cost of 521 if costs areroundedup or down to
the nearest integer, and a cost of 508 if truncatedinteger
costs are used. In instanceswith route durationconstraints,
using an insufficient precision may yield an infeasible
solution if travel times are proportionalto distances. For a
furtherdiscussion on the bias introducedby rounding in
VRPs, see Mole.27
Another issue related to accuracy is consistency. As a
rule, users will prefera heuristicthat performswell all the
time ratherthan one that may performeven bettermost of
the time but very poorly on other occasions, and may even
produce solutions easily perfectible by visual inspection.
Such solutions are enough to discreditthe algorithm.
Finally,users will often preferan algorithmthat produces
a good solutionat an early stage, and then displayssolutions
of increasingquality throughoutthe execution to an algorithmthatcomes up with only a final answer,possibly aftera
long computingtime. This gives users a betterfeel of how
much additionaleffort is worthinvestinggiven the evolution
rate of the solution value.

Four attributes of good VRP heuristics

Speed

Vehicle routingheuristics,as are most heuristics,are usually

measuredagainst two criteria:accuracy and speed. In our
opinion simplicityandflexibilityare also essential attributes
of good heuristics.We now elaborateon these four criteria.

Justhow importantis computationspeed in vehicle routing?

It all dependson the planninglevel at which the problemis
solved and on the degree of accuracy required. At one
extreme, real-time applications such as express courier
pickup and delivery28or ambulanceredeployment29require
fast, sometimes almost instantaneous,action. For example,
Gendreauet a129describethe crucialrole played by parallel
computing in a setting where an ambulance relocation
strategymust be determinedevery threeminuteson average.
At the otherextreme,in long termplanningdecisions made
every severalmonths, such as fleet sizing, it makes sense to
invest severalhoursor even severaldays of computingtime,
particularlyif large sums of money are at stake. Most
applications fall somewhere between these two extremes.
It does not seem unreasonableto invest ten or twenty
minutes of computingtime on a routingproblemthat must
be solved daily. Interactivesystems must of course react

Accuracy

Accuracy measures the degree of departureof a heuristic

solution value from the optimal value. Since optima and
sharplowerboundsare usually unavailablein the case of the
VRP, most comparisonshave to be made with best known
values. As pointed out by Barr et al,24 analysing heuristic
results is fraught with difficulties. Authors often report
results obtained for the best combination of algorithmic
parameters,or for the best of several runs with different
startingsolutions. Not all authorsuse the same roundingor
truncatingconventionswhich, as emphasizedby Gendreau

Vol.
Research
514 Journal
oftheOperational
53,No.5
Society

much more quickly. The issue of speed is not always

properly put into perspective. For example, would not
most practitionersprefer a VRP heuristic that is 2.38%
accurateand runs in 3.48 min to one that is 5.85% accurate
but requiresonly 0.26 min of computingtime? (These are
actual statistics of the 2-petal and 1-petal algorithms
describedby Renaudet al.30)
Accuratereportingof computingtime is anotherdelicate
issue in the scientific literature,particularlyin the case of
multipleruns or if parallelcomputingis used. As in the area
of accuracy,strict standardsare not consistenlyenforcedby
VRP researcherswhen it comes to assessing the speed of
algorithms.
Simplicity
SeveralVRP heuristicsare rarelyimplementedbecause they
arejust too complicatedto understandand to code. While it
is unrealisticto expect scientificarticlesto providea minute
descriptionof every algorithmicdetail, sufficient information should be provided to enable a reasonably skilled
programmerto come up with a working code. In addition,
heuristics should be reasonablyrobust to ensure that they
work properly,even if not every single detail is implemented. Many algorithmic descriptions fail on the count of
providingtoo much or insufficientdetail. One reason why
the Clarke and Wright16algorithm is so popular among
practitionersis that its basic principleis trivialto understand
and easy to code. (Amazingly, the original description is
rather clumsy and would probably fall short of today's
publishing standards.)Simple codes, preferablyshort and
self-contained, stand a better chance of being adopted,
although a minimum of complexity is to be expected for
good results.
Algorithmsthat containtoo many parametersare difficult
to understandand unlikely to be used. This problem is
prevalentin most metaheuristicsdevelopedover the past ten
years. In their quest for ever better solutions, researchers
have increasedthe numberof parameterscontainedin their
algorithms far beyond what can be deemed reasonable,
particularlyin view of the fact that relativelyfew instances
are used in the tests. This problemwas recently raised by
Golden et al.31 Not only should the numberof algorithmic
parametersbe limited,but these should also make sense to
the end-user.Thus a parametercontrolling the number of
consecutiveiterationswithoutimprovementin a local search
heuristic is easy to understand,whereas a bound on the
number of executions of an internal procedure is meaningless to most people.
There are two easy ways around the proliferation of
parameters.One is to set them once and for all at some
meaningfulvalue, especially if tests show thatthe algorithm
is rather insensitive to a particular parameter choice.
Another possibility is to make use of parametersthat selfadjustduringthe course of the algorithm.5'25

Flexibility
A good VRP heuristicshould be flexible enough to accomodate the various side constraintsencounteredin a majority
of real-life applications.While most of the VRP literature
focuses on capacityand sometimesroute length constraints,
it is often clear how changes can be made to deal with
additionalconstraints,but this is not always possible, and
performancecan also deterioratesignificantly as a result.
Our experiences525suggests that an efficient way of handling

side constraintsin a local search process is throughthe use

of two objectives.The first,F(x), computesthe routingcost
of solution x. The second, F'(x), is the sum of F(x) and
weighted penalty terms associated with violations of each
side constraint. For example, if Q(x) and D(x) are the
capacity and route durationviolations associatedwith solution x, then F'(x) would be defined as F(x) + aQ(x)+
pD(x), where cxand fi are positive self-adjustingpenalty
parameters.Initially set equal to 1, these parametersare
periodically increased or decreased throughoutthe search
accordingto whetherprevious solutions were infeasible or
feasible. This way of proceedingmeans that the searchwill
probably evolve through a mix of feasible and infeasible
solutions,thus reducingthe probabilityof becomingtrapped
in a local minimum.Anotheralgorithmicadvantageof this
device is that the search can operate with relatively simple
moves, such as removinga customerfrom its currentroute
and insertingit in a differentroute.When feasibilitymust be
maintainedat all cost, such simple moves tend not to work if
side constraints are tight, and more complex and time
consuming operations must then be envisaged, such as
ejection chains.32'33So in a sense, algorithmicflexibility is
in part achieved throughsimplicity of design.
Classical heuristics
Several heuristics have been devised for the VRP,12only
some of which are sufficentlywell knownto be trulyviewed
as 'classical'. For the sake of parsimonywe concentrateon
three of the best known heuristics:the Clarke and Wright
algorithm,the sweep algorithmand the Fisherand Jaikumar
algorithm. These were selected partly because of their
popularity,and also because they operateon vastly different
principles. Some extensions of these methods are also
discussed in passing.
The Clarkeand Wrightsavings heuristic
The Clarkeand Wright(CW)16heuristicis one of the best
known and remains widely used in practice to this day,
despite some of its shortcomings.It is based on the notion of
saving. Initially,a feasible solution consists of n back and
forthroutesbetweenthe depot and a customer.At any given
iteration,two routes (v, ... vi, v0) and (v0, vj, .. ,v0) are
merged into a single route (v0, ... , vi, v ...., v0) whenever

J-FCordeau
etal-Guide
tovehicle
heuristics
515
routing
this is feasible, thus generating a saving si = cio + c0J- Ci.

results in a sharp deterioration in solution quality. This

can be explained by the fact that the algorithm is based on
a greedy principle and contains no mechanism to undo early

sequential version keeps expanding the same route until

unsatisfactoryroute merges. Solomon36reportsa variantof

the CW algorithmin which savings are adaptedto handle
time windows, but results are disappointing.
Several improvementsto the CW algorithm have been

In the parallel version of the algorithm,the merge yielding

the largest saving is always implemented, whereas the
this is no longer feasible. In practicethe parallelversion is
much better. It is common to apply a 3-opt34 post-optimiza-

tion step to the final solution.

This algorithmscores very high on simplicity and speed.
It containsno parametersand is easy to code. In our tests on
the CMT benchmarkinstances,it typically ranwithin 0.12 s
on a Sun Ultrasparc10 workstation(440 MHz), withoutthe
3-opt step. Executing a 3-opt post-optimization step
increased the average running time to only 0.13 s. This
algorithm obtains a medium score on accuracy.The best
implementation (parallel version followed by 3-opt)
produced an average deviation of 6.71% from the best
known solution values identified by Taillard35and Rochat
and Taillard22(see Table 1). In addition,severalresearchers
have observed that the solution is sometimes of rather poor
quality and often contains at least one rather circumferential

route. The lack of flexibility is probablythe worst featureof

this algorithm.While additionalconstraintscan, in principle, be incorporatedin the CW algorithm, this usually
Table I
Clarkeand Wrighta
Instance n

Typed Value Secondse

1
50 C
2
75 C
3
100 C
4
150 C
5
199 C
6
50 C,D
7
75 C,D
8
100 C,D
9
150 C,D
10
199 C,D
11
120 C
12
100 C
13
120 C,D
14
100 C,D
Averagedeviation
from best and
time

578.56
888.04
878.70
1128.24
1386.84
616.66
974.79
968.73
1284.63
1521.94
1048.53
824.42
1587.93
868.50
6.71%

0.03
0.05
0.10
0.21
0.32
0.05
0.06
0.08
0.17
0.33
0.14
0.08
0.12
0.08
0.13

proposed. Gaskell37 and Yellow38 have suggested using

generalized savings of the form s, = cio + co - ACijto
help produce more compact routes, where A is a positive

parameter.Other enhancementsare related to the use of

to
sophisticated data structuresand sorting strategies39'40
better handle the savings. We believe, however,that, given
the present level of computertechnology and the very fast
running time of the CW algorithm on medium size
instances, the latter improvements are fast becoming
irrelevant.
Another stream of research on the CW algorithm has
concentrated on optimising the route merging process
through the use of a matching algorithm. Early results
based on this idea were produced by Desrochers and
Verhoog41and Altinkemer and Gavish.42 The best and
most recent implementationis due to Wark and Holt43

Comparisonof five classical heuristicsfor the VRP

Two-matchingb
Value

Secondsf

524.6k
835.8
830.7
1038.5
1321.3
555.4
911.8
878.0
1176.5
1418.3
1043.4
819.6
1548.3
866.4
0.63%

1200
3000
8700
17100
28800
1800
2700
9900
20700
32100
16500
5700
30600
8400
13371.42

SweepC

l-Petalc

Value Secondsg Value Secondsg

531.90
884.20
846.34
1075.38
1396.05
560.08
965.51
883.56
1220.71
1526.64
1265.65
919.51
1785.30
911.81
7.09%

7.2
10.2
70.8
151.8
216.0
9.6
11.4
88.2
180.0
294.6
211.2
38.4
134.4
5.1
105.6

531.90
885.02
836.34
1070.50
1406.84
560.08
968.89
877.80
1220.20
1515.95
1252.84
824.77
1173.69
894.77
5.85%

6.0
4.2
19.2
24.6
24.6
5.4
4.2
15.0
15.6
21.0
36.6
12.6
15.6
10.2
15.6

2-Petalc
Value Secondsg
524.61
45.6
854.09
31.2
830.04
230.4
1054.62
355.8
1354.23
372.6
560.08
33.6
922.75
25.8
877.29
174.6
1194.51
214.8
1470.31
311.4
1109.14
702.0
824.77
126.6
1585.20
198.6
885.87
101.4
2.38% 208.8

Best
524.61hj

835.26h
826.14h
1028.42h

1291.45'
555.43h
909.68h
865.94h

1162.55h
1395.85'
1042.11h
819.56h

1541.14h
866.37h

aImplementedby Laporte and Semet.'2 Verifiableresults (detailed solutions are available in articles or on web sites). Computations
performedby roundingcosts seven digits after the decimal point.
bImplementedby Warkand Holt.43Verifiableresults, best of five runs. Computationsperformedby roundingcosts two digits after the
decimal point.43
by Renaudet al.30 Verifiableresults. Computationsperformedby roundingcosts twelve digits after the decimal point.66
CImplemented
dC:capacity restriction,D: route durationrestriction.
eSun Ultrasparc10 workstations(42 Mflops).
fSun 4/630 MP Totaltime for five runs.
gSun Sparcstation2 (210.5 Mips, 4.2 Mflops).
hTaillard.35
Verifiableresults.
'Rochatand Taillard.22
jOptimalsolution value.
kBold numberscorrespondto best known values.

oftheOperational
Vol.
516 Journal
Research
53,No.5
Society

who often obtainedsignificantimprovementsover the original CW implementation, but at the expense of much
increased computing times. Their results, presented in
Table 1, correspondto the best of five runs, each requiring
on the average between 4 and 107min on a Sun4/630MP.
On the whole, this algorithmicenhancementremoves from
the CW heuristic two of its best features (speed and
simplicity), is difficult to implement and does nothing to
redress the lack of flexibility of the original algorithm.
While the repeated applicationof a matching based algorithmyields increasedaccuracy(it producesaveragedeviation of 0.63% on the CMT benchmark instances), this
technique has in our opinion limited potential because of
its complexity and low flexibility level.
The sweep algorithm
The sweep algorithmis generally attributedto Gillett and
Miller17althoughits principlecan be tracedback to Wren44
and Wren and Holliday.45It applies to planarinstances of
the VRP.Feasibleroutesare createdby rotatinga ray centred
at the depot and graduallyincluding customersin a vehicle
routeuntil the capacityor route length constraintis attained.
A new route is then initiated and the process is repeated
until the entire plane has been swept. A 3-opt step is then
typically applied.On the CMT instances,the Renaudet al30
implementationof this algorithm has yielded an average
deviation of 7.09% from the best known solution values.
Averagecomputingtimes of 105.6 s were obtainedon a Sun
Sparcstation2 (210.5 Mips, 4.2 Mflops) (see Table 1). This
algorithmscores high on simplicity,but does not seem to be
superiorto CW both in terms of accuracyand speed. It is
also ratherinflexible.Again, the greedy natureof the sweep
mechanism makes it difficult to accomodate extra
constraints and the fact that the algorithm assumes a
planarstructureseverelylimits its applicability.In particular,
the algorithmis not well suited to instances defined in an
urbansetting with a grid street layout.
A number of heuristics generate feasible vehicle routes
(sometimes called petals in this context) and determinea
best combinationthroughthe solution of a set partitioning
problem. Prime examples of this approachare the 1-petal
algorithmof Fosterand Ryan46and Ryan et a147 and the 2petal heuristicof Renaudet al30where, in additionto single
routes, double vehicle routes are also generated. These
extensions provide accuracy gains with respect to the
sweep algorithm. In terms of speed, improvements can
also be made since the 3-opt step is no longer applied (see
Table 1). These extensions do not get full marks on
simplicity since generatinga large pool of petals (especially
2-petals) can be cumbersome, and a set partitioningstep
must also be executed. As far as flexibility is concerned,
petal algorithmscan be made to accomodatea wide variety
of constraintsbut this can come at the expense of simplicity.
In fact, this type of algorithmcan be viewed as a truncated

version of column generationwhich is known to produce

high quality results on tightly constrainedVRPs with time
windows,48but again simplicity is sacrificed.
The Fisher and Jaikumaralgorithm
The Fisherand Jaikumar'8algorithmis a two-phaseprocess
in which feasible clusters of customersare first createdby
solving a generalized assignment problem (GAP), and a
vehicle route is determinedon each cluster by means of a
travellingsalesmanproblem(TSP) algorithm.To formulate
the GAP, it is necessary to first determinea seed for each
route from which customer distances are computed. Since
the GAP is NP-hard,it is usually solved by means of a
Lagrangianrelaxationtechnique. While good results were
reportedfor this algorithmin the early 1980s we have some
misgivings about its performance. The original article18
provides integer solutions values without providing the
rounding or truncatingrule, and the solutions cannot be
verified, which makes the assessment of the algorithm
difficult. Our own computationalexperimentssuggest that
the Fisherand Jaikumaralgorithmis not simple to program
and its speed is highly relatedto the choice of seeds and to
the implementationof Lagrangianprocess. We found that if
seeds are selected as in Fisher et a149 and the Lagrangian
steps programmedas recommendedby Fisher et al50 then
convergenceis often poor, severaltrialsmay be necessaryto
reach a satisfactorysolution, and even then accuracycan be
low. We did not, in general, succeed in obtaining results
close to those of Fisher and Jaikumar.18Flexibility is also
problematic. While it is, in principle, simple to handle
capacity constraints, the original reference contains no
indication regarding the treatment of route duration
constraints, although six instances involving such
constraintsare reportedlysolved. We see no easy mechanism by which to incorporatethese or other side constraints
within the GAP algorithm.Bramel and Simchi-Levi51have
optimized the choice of seeds in the Fisher and Jaikumar
algorithmby solving a capacitatedlocation problem.Their
resultson the seven CMT instancescontainingonly capacity
constraints show a significant average deviation (3.29%)
from the best known results (see Table 2).
Metaheuristics
Comparedwith classical heuristics,metaheuristicsperform
a much more thoroughsearch of the solution space, allowing inferior and sometimes infeasible moves, as well as
recombinationsof solutionsto createnew ones. This areaof
researchhas experienceda formidablegrowthover the past
ten years and has produced some highly effective and
flexible VRP heuristics.13It is fair to say, however, that
the gains in solution quality obtained with these modem
heuristicshave often been made at the expense of speed and
simplicity,althoughthis is not the case of some of the more

J-FCordeau
etal-Guide
tovehicle
heuristics
517
routing
Table 2 The Fisher and Jaikumaralgorithmand the Bramel and Simchi-Levienhancement
Fisher and Jaikumara
Instance

Typec

Value

Secondsd

1
2
3
4
5
6
7
8

50
75
100
150
199
50
75
100

C
C
C
C
C
C,D
C,D
C,D

524
857
833
1014
1420
560
916
885

9.3
12.0
17.7
33.6
40.1
15.2
20.6
52.2

9
10

150
199

C,D
C,D

1230
1518

121.3
136.6

11
12

100
120

C
C

13

100

C,D

14
120
C,D
Averagedeviationfrom best and time

824

6.4

848

6.3
47.13

Location-basedheuristicb
Value
524.6f
848.2
832.2
1088.6
1461.2

Secondse
68
406
400
2552
4142

Best
524.61g'
835.26g
826.14g
1028.42g
1291.45h

555.43g
909.68g
865.94g
1162.55g
1395.85h

1051.5
826.1

1303
400
-

3.29%

1042.11g
819.56g
1541.14

-866.37g
1324.43

Unverifiableresults. The roundingor truncatingrule is not reported.

almplementedby Fisher and Jaikumar.18
bImplementedby Brameland Simchi-Levi.5'Verifiableresults.Computationsperformedwith double-precisionfloatingpoint arithmetic.67
CC:capacity restrictions,D: route durationrestrictions.
dDEC-10computer.
eRS6000, Model 550.
fBold numberscorrespondto best known values.
gTaillard.35
hRochatand Taillard.22
'Optimalsolution value.26

recent implementations. The past decade has been very rich

in algorithmic design and testing. Some promising ideas did
not translate into equally good algorithms while others have
withstood the test of time.
Tabu search (TS) clearly stands out as the best metaheuristic for the VRP. By and large, the best TS implementations
for the VRP dominate other search processes such as
simulated52 and deterministic53'54 annealing, genetic
search,55 ant systems56 and neural networks57 (see, eg,
Gendreau et all3). The idea behind TS is to perform a
local search by moving, at iteration t, from a solution xt to
the best solution xt+l in its neighbourhood. Since moving
from xt to x,t+ may cause the objective function to deteriorate, an anti-cycling mechanism is put in place, namely any
solution possessing some attribute of x, is declared tabu, or
forbidden, for a number of iterations. So, in effect, the best
neighbour xt+l of x, is only selected if it is non-tabu or if it
is clear cycling will not occur. Several additional mechanisms, such as diversification and intensification, have been
implemented by a variety of researchers. Of course, not all
TS heuristics for the VRP have been equally successful. The
first known implementation, by Willard,58 did not make use
of a sufficiently powerful neighbourhood structure to allow
the identification of high quality solutions. In contrast,
several implementations that followed contained too many
devices and user-controlled parameters. There has been a
tendency in recent years toward leaner implementations. In

what follows, we present some of the best available TS

heuristics for the VRP.
Taburoute
With respect to earlier TS implementations, the Taburoute
heuristic of Gendreau et a125 is rather involved and contains
several innovative features. The neighbourhood of xt is the
set of all solutions reachable from xt by removing a vertex v
from its current route r and inserting it in another route s
containing one of its closest neighbours by means of a
generalized insertion (GENI) procedure.59 Reinserting v in
route r is then declared tabu for 0 iterations, where 0 is
randomly drawn from the interval [5, 10], as suggested by
Taillard60 in the context of the quadratic assignment
problem. Intermediate infeasible solutions are considered
and penalized through the use of self-adjusting penalty
parameters, as explained earlier. Periodic reoptimisations
of the individual vehicle routes are performed by means of
the GENIUS heuristic59 for the TSP. A continuous diversification strategy is also applied to penalise frequently moved
vertices. This is done by adding to the objective function a
term proportional to the relative past movement frequency
of the vertex currently being considered. False starts are also
employed: this allows a limited search starting from several
initial solutions, and a full search is then performed using
the most promising starting point. As is commonly done, the

518 Journal
oftheOperational
Research
Vol.
53,No.5
Society

procedureends when the objective has not improvedfor a

numberof consecutive iterations.
On the CMT test problems,Taburoutehas producedhigh
qualitysolutions:the averagedeviationfromthe best known
values is 0.86% and five best known solutions were
produced. Running times can be described as good but
not excellent. Solutiontimes vary between6 and 100 min on
Silicon Graphicsworkstation(36 MHz, 5.7 Mflops). Taburouteprobablyscores ratherlow on simplicity,partlydue to
its large numberof user-controlledparameters(nine in all)
and to its use of the outside TSP heuristicGENIUSwhich is
not that easy to reproduce.Flexibility,however,is relatively
high given that additionalconstraintscan easily be incorporatedthroughthe penalty mechanism.
The Taillardtabu search algorithm
The Taillard35TS implementationwas developed at the
same time as Taburoute.It also uses randomtabu durations
and continuous diversification.Where it differs most from
Taburoute is in the neighbourhood structure. Standard
vertex insertionsand exchanges are used instead of GENI.
Periodicroutereoptimizationsareperformedby means of an
exact TSP algorithm. To help speed up computations,
Taillardpartitions the problem into several subproblems,
each of which is solved independentlyon a parallelprocessor. In the case of planar problems, the decomposition
process uses concentric rings carved into sectors centred
at the depot. For non-planarproblems a different decomposition methodbased on computationof shortestspanning
arborescencesis used. The boundariesof the subproblems
are redefineddynamically.
Taillard'salgorithmis one of the best availablein termsof
accuracy.It has identifiedtwelve of the fourteenbest known
resultson the CMT instances.Computationtimes needed to
obtain best solutions are not reportedand are difficult to
establishwith any degree of accuracy.The algorithmis, in
one respect at least, simpler than Taburoutesince it uses
standardinsertions,but managingthe dynamic decomposition process as well as the parallelimplementationadds to
its complexity. One would expect this algorithmto handle
additional side contraintsreasonably well because of the
combinationof insertionand exchangemoves used to define
neighboursolutions. We believe, however,that resortingto
simple insertionsand a penalized objective function should
offer more flexibility.
The adaptive memoryprocedure of Rochat and Taillard
The concept of adaptivememory,developed by Rochat and
Taillard,22is probablyone of the most powerful ideas put
forwardin the area of metaheuristicsin recent years. An
adaptivememory is a pool of good solutionsproducedby a
heuristic,which is dynamicallyupdatedby addingto it new
high qualityelementsand removingfrom it some of its least

interestingelements.New elements are generatedby recombining good solutions from the pool. The memory update
process can therefore be viewed as a form of population
searchand as a generalizationof genetic search.Rochatand
Taillard22have developed an adaptivememory mechanism
for the capacityand routedurationconstrainedVRP and for
the VRP with time windows, based on the earlier TS
algorithmsof Taillard35and Rochat and Semet.61Given a
pool of good VRP solutions, new solutions are obtainedby
extractinghigh qualityvehicle routes, where routes belonging to bettersolutionsaregiven a higherprobabilityof being
selected. While extractingvehicle routescare is takennot to
include those that containalreadycoveredcustomers.Eventually this process will stop with a set of selected routesand
some unroutedcustomers.A new solution is then reconstituted from these routes and unroutedcustomersusing TS,
and included in the pool if it is of sufficient quality. Two
new best VRP solutions were identified through this
process.
The success of an adaptive memory procedure is
obviously linked to the capacity of the underlying search
processto generatea pool of high qualitysolutions,which is
certainlythe case of Taillard's35TS algorithm.It may not
work so well if an unsophisticatedheuristicwas employedto
generate the individual solutions. The use of an adaptive
memorycertainlyaddsto computationtime, as illustratedby
some of the results obtained by Rochat and Taillard.22
Coding an adaptivememory procedurerequiresa minimum
of computing skills but is not overly complicated. The
concept is highly flexible since it can be used in conjunction
with othertypes of heuristics,not only TS, and it can easily
be adaptedto other contexts. For example, Bozkaya et al62
reportan applicationin the area of political districting.
Thegranular tabu search algorithmof Tothand Vigo
The idea behind granulartabu search (GTS)63is to remove
from the graphunpromisingarcs or edges that have only a
small likelihood of belonging to an optimal solution. Toth
and Vigo63 eliminate all edges whose cost exceeds a
granularity threshold v = Pc, where /f is a sparsification
parameter, and c is the average cost of an edge in a good
solution generatedwith a fast heuristic.If / is chosen in the
interval [1.0, 2.0], then only 10-20% of the original edges
tend to remain.The value of this parameteris dynamically
updatedthroughoutthe searchprocess. In their implementation, Toth and Vigo work on the restricted edge set
E(v) = {(vi, vj) E E: cij < v} U I, where I is a set of impor-

tant edges such as those incident to the depot and those

belonging to high quality solutions. The searchmechanism
implementedby Tothand Vigo uses Taburouteas a subroutine, but tends to produce better results faster. Because it
works on a sparse graph, GTS can quickly perform an
extensive search of the solution space. It also produces
high quality solutions. The concept of granularity is

tovehicle
heuristics
519
J-FCordeau
eta-Guide
routing

relativelyeasy to implementonce a good underlyingsearch

algorithmis available.
The unified tabu search algorithmof Cordeau et al
The unified tabu search algorithm (UTSA) was initially
designed by Cordeau et al for the periodic VRP and the
multi-depotVRP,and laterslightly modifiedand extendedto
the single depot, multi-depotand periodic VRP with time
windows by Cordeau et al.64 The algorithm shares some
featureswith Taburoute,namely a penalised objective function F' with self-adjustingcoefficients to allow the exploration of intermediateinfeasible solutions, and the use of
continuousdiversification.However,fixed length tabu durations are used, and only one initial solution is generated.As
in Semet and Taillard,65
the tabu mechanismoperateson an
attributeset B(x) associated with solution x. More specifically, B(x) = {(i, k): vertex vi is visited by vehicle k
in solutionx}. Neighbour solutions are obtainedby removing an attribute(i, k) from B(x) and replacingit with (i, k'),
where k' : k. The insertion of vi into route k' is then
performedso as to minimize the penalized objective funcTable 3
Taburoutea
Instance
1
2
3
4
5
6
7
8

9
10

Type

50
C
75
C
100 C
150 C
199 C
50 C,D
75 C,D
100 C,D

Value

Minutese

tion F, and attribute(i, k) is declaredtabu for a set number

of iterations.
The algorithmwas highly successful on a varietyof VRPs
including the classical VRP with capacity and distance
restrictions (see Table 3), the periodic and multi-depot
VRPs,5 the VRP with site dependencies,4and the single
and multi-depotVRP with time windows.64On the classical
VRP, it yields solution values within 0.69% of the best
known. The averagecomputingtime is 13.75min on a Sun
Ultrasparc10 workstation(440 MHz). On all otherproblems
it often identifiednew best known solutions and has proved
superiorto several competing methods.
The UTSA ranks high on accuracy and performsrather
well on speed, althoughit cannotcompeteon this dimension
with very fast heuristicssuch as CW.Since all but one of its
parametersare the same for all applications,this heuristic
has in effect only one parameter.Moreover,since it operates
according to a very simple mechanism, it is probablythe
simplest of all TS implementationsfor the VRP, but still it
does not outperformCW in terms of simplicity. Finally,
UTSA obtains excellent markson flexibility.To our knowledge, this is the only VRP heuristic that can effectively

Comparisonof five metaheuristicsfor the VRP

Taillardb Adaptivememotyb

Granulartabu
searchc

Value

Value

Value

Minutesf

Unifiedtabu search
algorithmd
Value

Minutesg

Best

524.611
835.77
829.45
1036.16
1322.65
555.43
913.23
865.94

6.0
53.8
18.4
58.8
90.9
13.5
54.6
25.6

524.61
835.26
826.14
1028.42
1298.79
555.43
909.68
865.94

524.61
835.26
826.14
1028.42
1291.45
555.43
909.68
865.94

524.61
838.60
828.56
1033.21
1318.25
555.43
920.72
869.48

150 C,D 1177.76

199 C,D 1418.51

0.81
2.21
2.39
4.51
7.50
0.86
2.75
2.90

524.61
835.45
829.44
1038.44
1305.87
555.43
909.68
866.38

71.0
99.8

4.57
7.27
11.23
18.72
28.10
4.61
7.55
11.17

1162.55
1397.94

524.61'k
835.261
826.14'
1028.42'
1291.45i
555.43'
909.68'
865.94'

1162.55
1395.85

1173.12
1435.74

5.67
9.11

1171.81
1415.40

19.17
29.74

1162.55'
1395.85i
1042.11'

11

120

1073.47

22.2

1042.11

12

1042.11

100

1042.87

3.18

819.56

16.0

1074.13

819.56

14.15

819.56

819.56

1.10

13

120 C,D

819.56

10.99

1573.81

819.56'

59.2

1541.14

1541.14

1545.51

9.34

1568.91

65.7

14.53

866.37

1541.14'

866.37

866.37

1.41

866.53

10.65

866.37'

14

100 C,D

Average deviation
from best and time

866.37
0.86%

46.8

0.06%

0.00%

0.69%

3.84

0.69%

13.75

aStandardalgorithmwith one set of parameters.Verifiableresults. Computationsperformedby roundingcosts five digits afterthe decimal
point.
bVerifiableresults. Best of several runs. Computationtimes are not reported.Computationsperformedwith floating point arithemic.68
CTothand Vigo.63 Computationsperformedby roundingcosts three digits after the decimal point.69
dSolutionsobtainedby means of the Unified TabuSearchAlgorithmof Cordeauet al,5 withoutchangingthe parameters,and with 100 000
iterations.Computationsperformedwith double-precisionfloating point arithmetic.
eSilicon Graphicsworkstation(36 MHz, 5.7 Mflops).
fPentium200 MHz PC.
gSun Ultrasparc10 (440 MHz).
hC:capacity restrictions,D: route durationrestrictions.

'Taillard.35
JRochatandTaillard.22
solutionvalue.26
kOptimal
'Boldnumberscorrespond
to bestknownvalues.

oftheOperational
Research
Vol.
520 Journal
53,No.5
Society
Table 4 Assessment of some of the main VRP heuristics
Accuracy

Speed

Simplicity

Flexibility

Low
High
Low
Low
Medium
Difficult to assess
Medium

Very high
Very low
Medium-high
High
Medium
Medium
Low

Very high
Low
High
Medium
Medium
Low
Low

Low
Low
Low
Medium
Medium
Low
Low

Metaheuristics

Accuracy

Speed

Simplicity

Flexibility

Taburoute25
Taillard35

High
Very high

Medium
Lowb

Medium
Medium-low

High
High

Adaptive memory22'a

Very high

Lowb

Medium-low

High

High
High

Medium
Medium

Classical heuristics
Clarkeand Wright(CW)16
Two-matchingbased methods43
Sweep17
1-Petal47
2-Petal30
Fisher and Jaikumar(FJ)18
Locationbased FJ by Brameland
Simchi-Levi5l

Granulartabu search63a
Unified tabu search algorithm64

Medium
Medium

High
High

aTheseare not VRP algorithmsper se, but mechanismsthat can be incorporatedwithin anotheralgorithmto
improve its performance.Some of the ratings of these two features are therefore linked to those of the
underlyingalgorithm.They relateto the resultspresentedby Rochat and Taillard22and Toth and Vigo63for
specific implementations.
and Rochat and Taillard22do not providethe time requiredto obtain their best known solution.
bTaillard35
They do, however,reportthe time requiredto obtain a solution value within 1% or 5% of the best known.
These computationtimes can be ratherhigh.

handle so many different variants with a unique set of

algorithmicrules and a single user-controlledparameter.
We present in Table 3 comparativecomputationalresults
for the various VRP metaheuristicsjust described. Care
must be taken when comparing results since the Taillard
and adaptive memory solution values are the best over
several runs, whereas in the case of Taburoute,GTS and
UTSA, all reportedvalues correspondto a single run using
standardparametersettings. In the case of Taburouteand
UTSA, better solution values (sometimes correspondingto
the best) were obtainedby eitherchangingparametervalues
(for Taburoute)or by executing several runs with the same
parameters, starting from different randomly generated
initial solutions (for UTSA).
Summary and conclusion
In this guide to VRP heuristicswe have examined some of
the well-knownclassical methodsas well as some of the best
available metaheuristics. Contrary to a long established
custom of assessing heuristics against accuracy and speed
only, we have added what we perceive as two criticial
criteria:simplicityand flexibility.The resultsof our analysis
are summarizedin Table4.
None of the classical heuristics fares very well on
accuracy and flexibility. In this category, the celebrated
Clarkeand Wrightheuristichas at least the distinct advantage of being very quick and simple to implement. This
probably explains its ongoing popularity. However, in
contexts where vehicle routes must be planned over a long
horizon and large sums of money are at stake, it is worth

investing time and resources in a method that performs a

more extensive explorationof the search space.
Among the tabu search algorithms, there has been a
tendency in recentyears towardincreasedspeed and simplicity. The adaptive memory procedure and the granularity
principle are highly portable concepts that can be used
within almost any search process. Among stand-alone
heuristics, the unified tabu search algorithm scores very
well on all dimensions.

Acknowledgements-This paper was partly written while the third author

at the TechnicalUniversity
visited the Centrefor Trafficand Transportation
of Denmark.This researchwas partlyfundedby the CanadaResearchChair
in DistributionManagement,by CanadianNaturalSciences and Engineering Research Council (NSERC) under grants 227837-00, OGP00338816,
OGP0039682, and OGP0036662, and by the Fonds pour la Formationde
Chercheurset l'Aide a la Recherche(FCAR) undergrant2002-ER-73080.
This supportis gratefullyacknowledged.The authorsthankEdgarAlberto
Cabraland FrancoisGuertinfor their help with programming,and Oli BG
Madsen who provided some useful references. Thanks are due to two
referees for their valuablecomments.

References
1 Dantzig GB and Ramser JH (1959). The truck dispatching
problem.Mngt Sci 6: 80-91.
2 Gendreau M, Laporte G, Musaraganyi C and Taillard ED
(1999). A tabu search heuristic for the heterogeneous fleet
vehicle routing problem. Comput Opns Res 26: 1153-1173.
3 Mingozzi A, Giorgi S and Baldacci R (1999). An exact method

for the vehicle routingproblemwith backhauls.TransportSci

33:315-329.

J-FCordeau
eta-Guide
tovehicle
heuristics
521
routing
4 CordeauJ-F and LaporteG (2001). A tabu searchheuristicfor
the site dependentvehicle routingproblemwith time windows.
INFOR39: 292-298.
5 CordeauJ-F,GendreauM and LaporteG (1997). A tabu search
heuristic for the periodic and multi-depot vehicle routing
problems.Networks30: 105-119.
6 DrorM, LaporteG and TrudeauP (1994). Vehicle routingwith
split deliveries.Discr Appl Math 50: 239-254.
7 Golden BL and Assad AA (eds) (1988). Vehicle Routing:
Methodsand Studies.North-Holland:Amsterdam.
8 FisherML (1995). Vehicle routing.In: Ball MO, MagnantiTL,
MonmaCL andNemhauserGL (eds). NetworkRouting,Handbooks in OperationsResearchand ManagementScience, vol. 8.
North-Holland:Amsterdam,pp 1-33.
9 Desrosiers J, Dumas Y, Solomon MM and Soumis F (1995).
Time constrained routing and scheduling. In: Ball MO,
MagnantiTL, Monma CL and NemhauserGL (eds). Network
Routing:Handbooksin OperationsResearchand Management
Science, vol. 8. North-Holland:Amsterdam,pp 35-139.
10 CrainicTG and LaporteG (eds) (1998). Fleet Managementand
Logistics. Kluwer:Boston.
11 Toth P and Vigo D (eds) (2001). The Vehicle Routing
Problem. SIAM Monographs on Discrete Mathematics and
Applications.SIAM Publishing:Philadelphia,PA.
12 Laporte G and Semet F (2002). Classical heuristics for the
capacitatedVRP. In: Toth P and Vigo D (eds). The Vehicle
RoutingProblem.SIAMMonographson Discrete Mathematics
and Applications.SIAM Publishing:Philadelphia,PA, pp 109128.
13 GendreauM, LaporteG and Potvin J-Y (2002). Metaheuristics
for the capacitatedVRP. In: Toth P and Vigo D (eds). The
Vehicle Routing Problem. SIAM Monographs on Discrete
MathematicsandApplications.SIAMPublishing:Philadelphia,
PA, pp 129-154.
14 CordeauJ-F, Desaulniers G, Desrosiers J, Solomon MM and
Soumis F (2002). The VRP with time windows. In: TothP and
Vigo D (eds). The VehicleRoutingProblem.SIAMMonographs
on Discrete Mathematicsand Applications.SIAM Publishing:
Philadelphia,PA, pp 157-193.
15 TothP and Vigo D (1998). Exact solutionof the vehicle routing
problem. In: CrainicTG and LaporteG (eds). Fleet Management and Logistics. Kluwer:Boston, pp 1-31.
16 ClarkeG and WrightJR (1964). Schedulingof vehicles from a
centraldepotto a numberof deliverypoints. OpnsRes 12: 568581.
17 Gillett BE and Miller LR (1974). A heuristicalgorithmfor the
vehicle dispatchproblem.OpnsRes 22: 340-349.
18 Fisher ML and JaikumarR (1981). A generalizedassignment
heuristicfor vehicle routing.Networks11: 109-124.
19 KirkpatrickS, GellattCD Jr and Vecchi MP (1983). Optimization by simulatedannealing.Science 220: 671-680.
20 Glover F (1986). Future paths for integer programming
and links to artificialintelligence. ComputOpns Res 13: 533549.
21 Holland JH (1975). Adaptation in Natural and Artificial
Systems.Universityof MichiganPress:Ann Arbor,MI.
22 Rochat Yand TaillardED (1995). Probabilisticdiversification
and intensificationin local search for vehicle routing.J Heuristics 1: 147-167.
23 ChristofidesN, Mingozzi A and Toth P (1979). The vehicle
routing problem. In: ChristofidesN, Mingozzi A, Toth P and
Sandi C (eds). CombinatorialOptimization.Wiley, Chichester,
pp 315-338.
24 BarrRS, Golden BL, Kelly JP,Resende MGC and StewartWR
Jr (1995). Designing and reportingon computationalexperiments with heuristicmethods.J Heuristics 1: 9-32.

25 GendreauM, Hertz A and Laporte G (1994). A tabu search

heuristicfor the vehicle routingproblem.Mngt Sci 40: 12761290.
26 HadjiconstantinouEA, ChristofidesN and Mingozzi A (1995).
A new exact algorithmfor the vehicle routingproblembased on
q-paths and k-shortestpaths relaxations.In: GendreauM and
LaporteG (eds). Freight Transportation.Baltzer:Amsterdam,
pp 21-43.
27 Mole RH (1983). The curseof unintendedroundingerror:a case
from the vehicle schedulingliterature.J Opl Res Soc 34: 607613.
28 Ichoua S, GendreauM and Potvin J-Y (2000). Diversionissues
in real-timevehicle dispatching.TransportSci 34: 426-38.
29 GendreauM, LaporteG and Semet F (2001). A dynamicmodel
and parallel tabu search heuristic for real-time ambulance
relocation.Parallel Comput27: 1641-1653.
30 RenaudJ, Boctor FF and LaporteG (1996). An improvedpetal
heuristic for the vehicle routing problem.J Opl Res Soc 47:
329-336.
31 Golden BL, Wasil EA, Kelly JP and Chao I-M (1998). Metaheuristics in vehicle routing: In: Crainic TG and Laporte G
(eds). Fleet Managementand Logistics. Kluwer: Boston, pp
33-56.
32 Rego C and RoucairolC (1996). A paralleltabu search algorithmusing ejection chains for the vehicle routingproblem.In:
Osman IH and Kelly JP (eds). Meta-Heuristics:Theory and
Applications.Kluwer:Boston, pp 661-675.
33 Rego C (1998). A subpath ejection method for the vehicle
routingproblem.Mngt Sci 44: 1447-1459.
34 Lin S (1965). Computersolutions of the traveling salesman
problem.Bell Syst TechJ 44: 2245-2269.
35 TaillardED (1993). Paralleliterativesearchmethodsfor vehicle
routingproblems.Networks23: 661-673.
36 Solomon MM (1987). Algorithms for the vehicle routing and
scheduling problem with time window constraints.Opns Res
35: 254-265.
37 GaskellTJ (1967). Bases for vehicle fleet scheduling.OplRes Q
18: 281-295.
38 Yellow P (1970). A computationalmodificationto the savings
method of vehicle scheduling. Opl Res Q 21: 281-283.
39 Nelson MD, Nygard KE, Griffin JH and Shreve WE (1985).
Implementationtechniques for the vehicle routing problem.
ComputOpns Res 12: 273-283.
40 Paessens H (1988). The savings algorithm for the vehicle
routingproblem.Eur J Opl Res 34: 336-344.
41 Desrochers M and Verhoog TW (1989). A matching based
savings algorithmfor the vehicle routingproblem.Les Cahiers
du GERAD G-89-04, E1coledes HautesEtudes Commerciales:
Montreal.
42 Altinkemer K and Gavish B (1991). Parallel savings based
heuristicfor the delivery problem.OpnsRes 39: 456-469.
43 WarkP and Holt J (1994). A repeatedmatchingheuristicfor the
vehicle routingproblem.J Opl Res Soc 45: 1156-1167.
44 WrenA (1971). Computersin TransportPlanning and Operation. Ian Allan: London.
45 Wren A and Holliday A (1972). Computer scheduling of
vehicles from one or more depots to a number of delivery
points. Opl Res Q 23: 333-344.
46 Foster BA and Ryan DM (1976). An integer programming
approachto the vehicle scheduling problem. Opl Res Q 27:
307-384.
47 Ryan DM, HjorringC and Glover F (1993). Extension of the
petal method for vehicle routing.J Opl Res Soc 44: 289-296.
48 Desrosiers J, Soumis F and Desrochers M (1984). Routing
with time windows by column generation.Networks14: 545565.

522 Journal
oftheOperational
Research
Vol.
Society
53,No.5
49 FisherML, GreenfieldAJ, JaikumarR and LesterJT III (1982).
A computerizedvehicle routing application.Interfaces 12(4):
42-52.
50 Fisher ML, JaikumarR and Van WassenhoveLN (1986). A
multiplieradjustmentmethod for the generalized assignment
problem.Mngt Sci 32: 1095-1103.
51 Bramel JB and Simchi-Levi D (1995). A location based
heuristicfor generalroutingproblems.Opns Res 43: 649-660.
52 Osman IH (1993). Metastrategysimulatedannealing and tabu
searchalgorithmsfor the vehicle routingproblem.Annal Opns
Res 41: 421-451.
53 Dueck G and ScheuerT (1990). Thresholdaccepting:a general
purposeoptimizationalgorithm.J ComputPhys 90: 161-175.
54 Dueck G (1993). New optimizationheuristic:the great deluge
algorithmand the record-to-recordtravel.J ComputPhys 104:
86-92.
55 Van Breedam A (1996). An analysis of the effect of local
improvementoperators in genetic algorithms and simulated
annealing for the vehicle routing problem. RUCA Working
Paper96/14, Universityof Antwerp:Belgium.
56 BullnheimerB, HartlRFand StraussC (1999). An improvedant
system algorithmfor the vehicle routingproblem.Annal Opns
Res 89: 561-581.
57 Ghaziri H (1996). Supervision in the self-organizing feature
map: applicationto the vehicle routingproblem.In: OsmanIH
and Kelly JP (eds). Meta-Heuristics:Theoryand Applications.
Kluwer:Boston, pp 651-660.
58 Willard JAG (1989). Vehicle routing using r-optimal tabu
search, MSc dissertation,The ManagementSchool, Imperial
College: London.

59 GendreauM, HertzA and LaporteG (1992). New insertionand

postoptimization procedures for the traveling salesman
problem.Opns Res 40: 1086-1094.
60 Taillard ED (1991). Robust taboo search for the quadratic
assignmentproblem.Parallel Comput17: 433-445.
61 Rochat Y and Semet F (1994). A tabu search approach for
deliveringpet food and flour in Switzerland.J Opl Res Soc 45:
1233-1246.
62 Bozkaya B, Erkut E and Laporte G (2002). A tabu search
heuristicand adaptivememory procedurefor political districting. Eur J Opl Res, to be published.
63 Toth P and Vigo D (1998). The granulartabu search (and its
applicationto the vehicle routing problem). Technicalreport
OR/98/9, DEIS, Universitadi Bologna: Italy.
64 CordeauJ-F, LaporteG and MercierA (2001). A unified tabu
search heuristic for vehicle routing problems with time
windows. J Opl Res Soc 52: 928-936.
65 Semet F and Taillard ED (1993). Solving real-life vehicle
routing problems efficiently using tabu search. Annal Opns
Res 41: 469-481.
66 RenaudJ (2001). Privatecommunication.
67 Simchi-LeviD (2001). Privatecommunication.
68 TaillardED (2001). Privatecommunication.
69 TothP (2001). Privatecommunication.

Received June 2001;

2001
November
after one revision
accepted