Journal of the Operational Research Society (2003) 54, 1254–1262 r 2003 Operational Research Society Ltd.

d. All rights reserved. 0160-5682/03 $25.00

www.palgrave-journals.com/jors

A new hybrid genetic algorithm for the capacitated

vehicle routing problem
J Berger* and M Barkaoui
Defence Research and Development Canada - Valcartier, Val-Be´lair, Canada
Recently proved successful for variants of the vehicle routing problem (VRP) involving time windows, genetic
algorithms have not yet shown to compete or challenge current best search techniques in solving the classical
capacitated VRP. A new hybrid genetic algorithm to address the capacitated VRP is proposed. The basic scheme
consists in concurrently evolving two populations of solutions to minimize total travelled distance using genetic
operators combining variations of key concepts inspired from routing techniques and search strategies used for a time
variant of the problem to further provide search guidance while balancing intensification and diversification. Results
from a computational experiment over common benchmark problems report the proposed approach to be very
competitive with the best-known methods.
Journal of the Operational Research Society (2003) 54, 1254–1262. doi:10.1057/palgrave.jors.2601635

Keywords: vehicle routing problems; heuristics

Introduction combining selection, recombination and mutation opera-

In the classical vehicle routing problem (VRP),1 customers
with known demands and service time are visited by a
homogeneous fleet of vehicles with limited capacity and
initially located at a central depot. Routes are assumed to
start and end at the depot. The objective is to minimize total
travelled distance, such that each customer is serviced
exactly once (by a single vehicle), total load on any vehicle
associated with a given route does not exceed vehicle
capacity, and route duration combining travel and service
time is bounded to a preset limit.
A variety of algorithms including exact methods and
efficient heuristics have already been proposed for VRP. For
a survey on the capacitated VRP and variants see Toth and
Vigo.1 Toth presents both exact and heuristic methods
developed for the VRP and its main variants, focusing on
issues common to VRP. Overview of classical heuristics and
metaheuristics may be found in Toth and Vigo1 and in
Gendreau et al.2,3
Tabu search4,5 techniques and (hybrid) genetic algorithms
represent some of the most efficient metaheuristics to
address VRP and/or its variants. The basic idea in tabu
search is to allow the selection of worse solutions once a
local optimum has been reached. Different memory struc-
tures are then used to prevent repeating the same solutions
(cycling), and to diversify and intensify the search. Genetic
algorithms6–8 are adaptive heuristic search methods that
mimic evolution through natural selection. They work by
*Correspondence: J Berger, Defence Research and Development Canada -
Valcartier, 2459 Pie-XI Blvd North, Val-Be´lair, PQ, Canada G3J 1X5.
E-mail: jean.berger@drdc-rddc.gc.ca
tions. The selection pressure drives the population toward
better solutions, while recombination uses genes of selected
parents to produce offspring that will form the next
generation. Mutation is used to escape from local minima.
Hybrid genetic algorithms combine the above scheme with
heuristic methods to further improve solution quality. Tabu
search heuristics have proved so far the most successful
technique for the capacitated VRP.9,2 Alternatively, despite
its relative success reported for the travelling salesman
problem (see Gendreau et al2) and variants of the VRP
involving time windows,2,10–19 genetic algorithms have not
yet shown to compete with tabu search techniques in solving
the capacitated VRP. Limited work using genetic-based
techniques for the classical capacitated VRP reports
mitigated success so far. As recently proposed procedures
match the performance of well-known classical methods,20
others fail to report comparative performance with the best
well-known routing techniques, while sometime demonstrat-
ing prohibitive run-time to obtain modest solution qual-
ity.13,21 Genetic-based methods recently developed for VRP
interleaving local improvement procedures through critical
steps of the standard genetic algorithm tend to provide good
solutions but have not convincingly shown to our knowl-
edge, to compete or challenge the best-known methods. It is
nonetheless believed that genetic-based methods targeted to
the classical capacitated VRP have not yet been fully
exploited. Accordingly, we contend that some benefits might
be expected in capturing heuristic knowledge in genetic
operators explicitly.
In this paper, a competitive hybrid genetic algorithm
(HGA-VRP) to address the classical capacitated VRP is

proposed. It consists in concurrently evolving two popula-
tions of solutions subject to periodic migration in order to
minimize total travelled distance using genetic operators
combining variations of key concepts inspired from routing
techniques and search strategies used for a time variant of
the problem to further provide search guidance while
balancing intensification and diversification. Granular
neighbourhoods were also exploited to further reduce
computational cost. A computational experiment conducted
on common benchmark problems shows the proposed
hybrid genetic approach to be competitive with the best-
published methods.
The paper is outlined as follows. The main concepts of the
proposed hybrid genetic algorithm (HGA) are first intro-
duced. Then, the selection scheme, recombination and
mutation operators are presented. Concepts derived from
well-known heuristics such as large neighbourhood search,22
route neighbourhood-based two-stage metaheuristic23 and l-
interchange mechanism24 are briefly outlined. The next
section presents an improved version of the algorithm. It
consists in reducing computational cost by restricting
exploration of the search space using granular neighbour-
hoods.25 The results of a computational experiment to assess
the value of the proposed approach reporting comparative
performance to alternate methods are then presented.
Finally, some conclusions and future research directions
are briefly outlined in the last section.
Hybrid genetic approach
General description
The proposed HGA-VRP algorithm aims at computing
high-quality solutions to compete with the best-known
procedures under acceptable run-time conditions. It relies
on the synergy between evolutionary principles and heuristic
methods to further provide search guidance while maintain-
ing a suitable balance between solution exploitation and
exploration.
Emphasizing genetic diversity, our approach consists in
concurrently evolving two populations of solutions (Pop1,
Pop2), while exchanging a certain number of individuals
(migration) at the end of a new generation. Exclusively
formed of feasible solution individuals, populations are
evolved to minimize total travelled distance using genetic
operators based upon variations of known routing methods.
Whenever a new best solution emerges, a post-processing
procedure (RC_M) aimed at reordering customers is applied
to further improve its solution quality. The RC_M mutation
operator is introduced in a subsequent section. The
evolutionary process is repeated until a predefined stopping
condition is met.
The proposed steady-state genetic algorithm resorts to
overlapping populations to ensure population replacement
for Pop1 and Pop2. At first, new individuals are generated
J Berger and M Barkaoui—Genetic alogorithm for capacitated VRP 1255
and added to population Popp (p ¼ 1; 2). The process
continues until the overlapping population outnumbers the
initial population by np. Then, the np worst individuals are
eliminated to maintain population size using the following
individual evaluation:
Evali ¼ di = maxðdm ; di Þ ð1Þ
where di is the total travelled distance related to individual i
and dm is the average total travelled distance over the
individuals forming the initial populations.
The lower the evaluation value the better the individual
score (minimization problem). An elitist scheme is also
assumed, in which the fittest solution individual automati-
cally survives to the next generation.
A direct problem-specific representation is used in which
ordered lists of customers define solution routes. No explicit
solution encoding is required, making identical the pheno-
type and genotype of a solution individual. Genetic
operators are simply applied to a population of feasible
solutions rather than a population of encoded solutions.
The general algorithm is specified as follows:
Initialization
Repeat
p¼1
Repeat {evolve population Popp - new generation}
For j ¼ 1..np do
Select two parents from Popp
Generate a new solution Sj using
recombination and mutation operators
associated with Popp
Add Sj to Popp
end for
Remove from Popp the np worst individuals using
the evaluation function (1)
p¼pþ1
Until (all populations Popp have been visited)
if (new best solution) then apply RC_M on best
solution {cust. reordering}
Population migration {local best solutions exchange
across populations}
Until (convergence criteria or max number of
generations)
The initialization phase involves the generation of initial
populations Pop1 and Pop2 using a random procedure to
construct feasible solution individuals. Solutions are gener-
ated using a sequential insertion heuristic in which customers
are inserted in random order at randomly chosen insertion
positions within routes. This strategy is fast and simple while
ensuring unbiased solution generation. The migration
consists in exchanging local best individuals from one
population to another. Convergence is assumed to occur
either when solution quality fails to significantly improve
1256 Journal of the Operational Research Society Vol. 54, No. 12
over a consecutive number of generations or, after a
maximum number of generations.
The proposed technique is significantly different from the
algorithm presented by Berger et al12 in many respects. The
new procedure focuses on a single objective, deals with a
different constraint set, handles feasible solutions only, and
presents a simpler problem-solving method involving new
and more efficient operators.
Selection
The selection process consists in choosing two individuals
(parent solutions) within the population for mating pur-
poses. The selection procedure is stochastic and biased
toward the best solutions using a simple roulette-wheel
scheme.8 In this scheme, the probability to select an
individual is proportional to its fitness value. Individual
fitness for both populations Pop1 and Pop2 is computed as
follows:
fitnessi ¼ di
The notation is the same as in Equation (1). Better
individuals show a shorter total travelled distance (mini-
mization problem).
Genetic operators
Genetic operators incorporate and combine key feature
variations of efficient routing techniques successfully applied
for the VRP with time windows,1 providing search guidance
through promising regions of the solution space while
reducing computation time. Even at the cost of making
these specialized operators hardly reusable for other
problems, the potential gain in performance may be quite
noticeable. Procedure variations include Solomon’s inser-
tions heuristic I1,26 in which non-deterministic contributions
have been added to both the insertion cost function and the
customer selection process to further intensify exploration;
large neighbourhood search,22 adding non-determinism to
the customer reinsertion process and, making the customer
visit ordering scheme less myopic by using additional
information to better handle multiple insertion opportu-
nities; and the l-interchange mechanism24 in which customer
sets are exchanged over alternate neighbouring routes
exploiting a new route neighbourhood structure inspired
from the RNETS23 metaheuristic to further reduce search.
Details on recombination and mutation operators are given
in the next sections.
Recombination
A single recombination operator is considered, namely
IB_X(k). Its contribution is to build upon promising
segments of two parent solutions to generate a new
improved child solution using a variant of a well-known
customer insertion heuristic. The idea is to improve as much
as possible some routes from one parent using ‘‘good genes’’
(good customer sequences) from the other. The insertion-
based IB_X(k) crossover operator creates an offspring by
combining, one at a time, k routes (R1) of parent solution P1
with a subset of customers, formed by nearest-neighbour
routes (R2) in parent solution P2. Routes defining R1 are
determined using a randomly selected criterion out of three.
Accordingly, routes are chosen randomly following a
uniform probability distribution, or selected with a prob-
ability proportional either to the number of customers
characterizing a tour, or to the average distance separating
consecutive route members. Child route generation follows
an iterative process. The neighbourhood R2 associated to a
given route r1AR1 is first created. R2 includes the routes of
P2 whose centroid is located within a certain range of r1
(centroid). A route centroid corresponds to a virtual site
whose coordinates refer to the average position of its specific
routed customers. The range characterizing R2 is given by
ð2Þ average distance separating r1 from the routes defining P2. A
stochastic removal procedure is then carried out to remove
from r1, customers likely to be migrated to alternate routes.
Targeted customers are either selected according to waiting
times, distance separating them from their immediate
neighbours, or randomly. Then, using a modified insertion
heuristic inspired from Solomon26 a feasible child tour is
constructed, expanding the altered route r1 by inserting
customer visit candidates derived from the nearest-neigh-
bour routes R2 defined earlier. The proposed insertion
technique consists in adding a stochastic feature to the
standard customer insertion heuristic I1,26 by randomly
using cost function parameters and selecting the next
customer visit over the three best candidates with a bias
toward the best. Once the construction of the child route is
completed, and reinsertion is no longer possible, a new route
construction cycle is initiated. The overall process is repeated
for the k routes of R1. Finally, the child inherits the
remaining ‘‘diminished’’ routes (if any) of P1. If unvisited
customers still remain, additional routes are built using a
nearest-neighbour procedure.
The whole process is then reiterated to generate a second
child interchanging P1 and P2. IB_X(k) can be outlined as
follows:
Step 1 Select a set R1 of k routes from P1 to be visited.
Step 2
Repeat
2.1 Select randomly an unvisited route r1 A R1
(chromosome segment) from P1 (chromosome).
2.2 Select a subset of routes R2 from P2 located in
the neighbourhood of r1. Customers forming the
routes of R2 define a new set NC2.
2.3 Remove from r1 customers likely to be migrated
to alternate routes (large waiting times or

distance separating neighbour route members) or
randomly.
2.4 Build a child route solution: apply a stochastic
version of the insertion-based routing technique
I1 in which the modified route r1 is the initial
child route solution and customers to be
S
reinserted run over NC2 U. U refers to the set
of unrouted but already visited customers.
2.5 Inserted customers are then removed from the
remaining unvisited routes of R1. Update U.
Until (all routes r1AR1 have been visited).
Step 3 Inherit the remaining unvisited routes of P1 (if
any) while eliminating customers already
routed.
Step 4 For the remaining unrouted customers U (if
any), build new routes using the nearest-
neighbour procedure.
Mutation
A suite of four mutation operators is proposed, namely
LNSB_M(d), EE_M, IEE_M and RC_M(I). Each mutator
is briefly described next.
The LNSB_M(d) (large neighbourhood search based)
mutation operator relies on the concept of the large
neighbourhood search (LNS) method proposed by Shaw.22
The contribution of this operator consists in mutating a
solution individual by conducting a more extensive search
(larger scale) of the solution space, while directing and
confining exploration to most promising regions, and keeping
computation to an acceptable cost. LNS consists in exploring
the search space by repeatedly removing related customers
and reinserting them using constraint-based tree search
(constraint programming). Customer relatedness defines a
relationship linking two customers based upon specific
properties (eg proximity and/or identical route membership),
such that when both customers are considered simultaneously
for a visit, they can compete with each other for reinsertion
creating new opportunities for solution improvement. There-
fore, customers close to one another naturally offer inter-
change opportunities to improve solution quality. Similarly,
solution number of tours is more likely to decrease when
customers sharing route membership are removed all
together. As stated in Shaw,22 a set of related customers is
first removed. The reinsertion phase is then initiated.
The proposed customer reinsertion technique differs from
the procedure introduced by Shaw22 resorting to alternate
insertion cost functions and, customer visit ordering schemes
(variable ordering scheme), using new information to carry out
large neighbourhood search. Customer visit ordering deter-
mines the effective sequence of customers to be consecutively
visited while exploring the solution space (search tree expan-
sion). It is believed that exploiting additional information
J Berger and M Barkaoui—Genetic alogorithm for capacitated VRP 1257
For diversification purposes, two customer reinsertion methods
are proposed, one of them being randomly selected (50%
probability) on mutator invocation.
The first reinsertion method relies on the insertion cost
function prescribed by Solomon’s procedure I126 for the
VRP with time windows and, a rank-based customer visit
ordering scheme. Customer insertion cost is defined by the
sum of key contributions referring, respectively, to increased
travelled distance and delayed service time. As for customer
ordering, customers ({c}) are sorted (CustOrd) according to
a composite ranking, departing from the myopic scheme
originally proposed by Shaw. The ranking is defined as an
additive combination of two separate rankings, previously
achieved over best insertion costs (RankCost(c)) on the
one hand, and number of feasible insertion positions
(Rank|Pos|(c)) on the other hand:
CustOrd SortðRankCost ðcÞ þ RankjPosj ðcÞÞ ð3Þ
The smaller the insertion cost (short total distance, travelled
time) and the number of positions (opportunities), the better
(smaller) the ranking. The next customer to be visited within
the search process is selected according to the following
expression:
customer CustOrd½INTEGERðLrand D Þ ð4Þ
where L is the current number of customers to be inserted,
rand is a real number over the interval [0,1] (uniform random
number generator), and D is the parameter controlling
determinism. If D ¼ 1 then selection is purely random
(default: D ¼ 15).
Customer position selection (value ordering) is then based
on cheapest insertion cost.
The second reinsertion method involves features of the
successful insertion heuristic proposed by Liu and Shen23 for
the VRP with time windows, relying on the maximization of
a regret insertion cost function, which concurrently takes
into account multiple insertion opportunities (regret cost), to
determine customer visit ordering. The regret cost based
customer visit ordering scheme is specified as follows. In the
insertion procedure proposed by Liu and Shen,23 route
neighbourhoods associated to unvisited customers are
repeatedly examined for customer insertion. This new route
neighbourhood structure relates one or multiple routes to
individual customers. In our approach, customer route
neighbourhood is strictly bounded to two tours, comprising
routes whose distance separating their centroid from the
customer location is minimal. Each feasible customer
insertion opportunity is explored over its entire route
neighbourhood. The next customer visit is selected by
maximizing the so-called regret cost function that accounts
for multiple route insertion opportunities:
X
already available from computation is likely to guide search Regret Cost ¼ fCc ðrÞ  Cc ðr
Þg ð5Þ
toward promising regions of the solution space more efficiently. r2RNðcÞ
1258 Journal of the Operational Research Society Vol. 54, No. 12
where RNðcÞ is the route neighbourhood of customer c,
Cc ðrÞ is the minimum insertion cost of customer c within
route r (Liu and Shen23), and Cc ðr
Þ is the minimum RC (if any) are then inserted in S using
insertion cost of customer c over its route neighbourhood.
For both reinsertion methods, once a customer is selected,
search is carried out over its different insertion positions
(value ordering) based on insertion cost minimization,
exploiting limited discrepancy search27 as specified in Shaw22
(a discrepancy refers to a decision point where branching
runs against the value ordering heuristic). However,
search tree expansion is achieved using a non-constant
discrepancy factor d, selected randomly (uniform probability
distribution) over the set {1,2} to reduce search signifi-
cantly. Remaining unvisited customers (if any) are then
inserted in additional routes. LNSB_M(d) can be outlined as
follows:
Step 1 Eliminate from solution S a random number of
related customers (see Shaw22) forming a set RC.
Step 2 Perform customer reinsertion of the set RC.
Perform constraint-based limited discrepancy
search (Shaw22) calling the recursive procedure
Search (S,RC,d) in which the discrepancy factor
d is selected randomly over {1,2} and using one
of the following combination with a 50%
probability:
(a) insertion cost function proposed by
Solomon;26 rank-based customer visit
ordering scheme,
(b) insertion cost function proposed by Liu
and Shen;23 regret cost-based customer
visit ordering scheme.
Search (partial solution S, RC, discrepancy d)
{ Step 1 For each customer c in solution S do
If RC ¼ {} then update best computed solution
else
For each unrouted customer cARC do
Compute insertion costs for all feasible
candidate positions.
end for
If (at least one feasible candidate position
exists) then
Select a customer c using the customer
visit ordering scheme.
Expand tree search: visit feasible
insertion positions (POSc) for customer
c in order of increasing insertion cost
subject to a limited number of
discrepancies:
i¼0
For pos ¼ 1..|POSc| and while (ipd) do
Insert c in S in position pos
Search (S,RC-{c},d-i)
i¼iþ1
end for
else
Remaining unrouted customers from
additional routes.
Search(S,{},0)
}
Step 3 Return mutated (best computed) solution.
The EE_M (edge exchange) mutator focuses on inter-
route improvement. EE_M attempts to shift customers to
alternate routes as well as to exchange sets of customers
between two routes. The contribution of this operator
consists in evolving a solution individual performing limited
search and therefore, is complementary to LNSB_M(d). It
relies upon a cheap and simple local search technique
making use of a route neighbourhood structure to further
reduce solution space exploration and then, computational
cost. It is inspired from the l-interchange mechanism of
Osman,24 performing reinsertions of customer sets over two
neighbouring routes. In the proposed mutation procedure,
each customer is explored for reinsertion in its surrounding
route neighbourhood made up of two tours. Tours are being
selected such that the distance separating their centroid from
customer location is minimal. Customer exchanges take
place as soon as the solution improves, that is, we use a ‘‘first
admissible’’ improving solution strategy. Assuming the
notation (x, y) to describe the different sizes of customer
sets to be exchanged over two routes, the current operator
explores values running over the range (x ¼ 1, y ¼ 0,1,2). The
IEE_M (intra-route edge exchange) mutation operator is
similar to EE_M except that customer migration is restricted
to the same route. EE_M can be outlined as follows:
Define a route neighbourhood RN(c).
Step 2 Repeat
Explore a feasible permutation between c
(x ¼ 1) and alternate sequences of y (0,1,2)
consecutive customers from neighbouring
routes rARN(c).
Until (solution improves or all sequences y
over all rARN(c) have been visited)
Step 3 if (solution improves) then
Exchange sequences x and y in solution S.
end for
Step 4 Return mutated solution S.
The RC_M(I) (reorder customers) mutation operator is an
intensification procedure intended to reduce total travelled
distance of feasible solutions by reordering customers within
a route. The contribution of this operator is to seek solution
quality improvement whenever a new best solution is
computed. The procedure consists in repeatedly reconstruct-
ing a new tour using the sequential insertion heuristic I1 over
I different sets (eg I ¼ 20) of randomly generated parameter

values, returning the best solution generated shall an
improved one emerge. RC_M(I) can be outlined as follows:
Step 1 For each route r in solution S do
rmut ¼ r
Let U be the set of customers defining route r.
Step 2 For i ¼ 1..I
Select randomly parameters from the
insertion procedure I1.
Build a new route r 0 using I1 over the set U
starting with the farthest customer from the
depot.
if (r 0 improves rmut and U ¼ { }) then rmut ¼ r 0
end for
Step 3 Replace route r by rmut in solution S.
end for
Step 4 Return mutated solution S.
Improved hybrid genetic algorithm
Improvements have been brought to HGA-VRP genetic
operators to further reduce run-time computation. Based
upon granular neighbourhoods,25 HGA-VRP* consists in
limiting exploration of the search space while reinserting
unvisited customers in a partial solution by restricting
possible visits to the most likely insertion positions
characterizing good feasible solutions. For a given customer,
such a neighbourhood is mainly defined by surrounding
routed customers located within a certain range or threshold
distance d*. It also includes the depot and routed customers
whose connection with the given customer is part of the best
computed solutions (past and current) so far. Threshold
distance d* is proportional to the average distance separating
consecutive customers in the best computed solution:
d
¼ dmin =ðn þ rmin Þb ð6Þ Recombination:
where dmin is the total travelled distance of the best
computed solution, rmin is the number of routes within the
best computed solution, n is the number of customers, and b
is the positive sparsification parameter.25
b is mainly used to define various neighbourhoods as it
can be changed dynamically to emphasize solution intensi-
fication (decrease) or diversification (increase) when needed.
Updated on a generation basis, b is either decreased or
increased by a positive contribution Db depending on
whether best computed solution quality improves or not.
Running into the interval [bmin, bmax], b is initially set to its
minimal value. d* is therefore modified in accordance with
Equation (6).
The insertion procedures embedded in the operators
IB_X(k), LNSB_M(d) and RC_M(I) were then modified to
carry out search based upon the granular neighbourhoods
described above. As for the revisited EE_M and IEE_M
J Berger and M Barkaoui—Genetic alogorithm for capacitated VRP 1259
operators, customers subject to permutation can be swapped
only if the distance separating each other is less than d*.
Computational results
A computational experiment has been conducted to compare
the performance of the proposed algorithm with some of the
best techniques designed for VRP. The algorithm has been
tested on the well-known VRP benchmark proposed by
Christofides et al.28 For these instances, travel time
separating two customers corresponds to their relative
Euclidean distance. In conformity to the comparative study
published in Cordeau et al,9 the experiment consisted in
performing a single simulation run for each problem
instance and reporting on average performance. Then,
additional simulations were conducted for HGA-VRP over
a larger sample and results reported accordingly. HGA-VRP
and HGA-VRP* have been implemented in C þ þ , using
the GAlib genetic algorithm library of Wall29 and the
experiment carried out on a 400 MHz Pentium processor.
Solution convergence is assumed to occur when its quality
fails to improve by at least 1% over 20 consecutive
generations. The parameter values for the investigated
algorithm are described below.
In the LNSB_M(d) mutation operator, the number of
customers considered for elimination runs in the range [15,
21]. The discrepancy factor d is randomly chosen over {1,2}.
Parameter values for the proposed genetic operators are
defined as follows:
Population size: 15
Migration: 5
Population replacement:
Elitism
Population overlap per generation: n1 ¼ n2 ¼ 2
IB_X(k ¼ 2) (20%)
Mutation:
LNSB_M(d) (80%)
EE_M (50%), IEE_M (50%)
RC_M(I ¼ 20)—whenever a new best feasible
solution is found.
The migration parameter, a feature provided by GAlib, refers
to the number of (best) chromosomes exchanged between
populations after each generation. Because of limited
computational resources, parameter values were determined
empirically over a few intuitively selected combinations,
choosing the one that yielded the best average output.
Comparative performance is reported for some of the best-
known VRP methods, namely referred to as OS (Osman24),
GHL (Gendreau et al30), CGL (Cordeau et al31), TV (Toth
and Vigo25), WH (Wark and Holt32), RR (Rego and
Roucairol33), RT (Rochat and Taillard34), TA (Taillard35)
1260 Journal of the Operational Research Society Vol. 54, No. 12

and BB for HGA-VRP. The results are expressed in terms of
total travelled distance. Published competing methods with an
average performance gap exceeding about 1% (over all
instances) of the best-known result, and/or failing to specify
run-time and computational resource characteristics, or
reporting prohibitive run-time have been deliberately omitted
for comparison purposes. Additional results involving other
techniques including classical heuristics may nonetheless be
found in Cordeau et al.9
Computational results for all problem data sets are
summarized in Table 1, similarly to the study of Cordeau
et al.9 The first column describes the various instances and
their related size, whereas the second specifies total travelled
distance and run-time (in minutes). The following columns
refer to particular problem-solving methods. Best-known
results (bold values) are depicted in the last column
(Taillard35 and, Rochat and Taillard34 for instances 5 and
10). The last row refers to average run-time and performance
deviation from the best-known solutions over all problem
instances. Related computer platforms include VAX 8600
for OS, Silicon Graphics 36 MHz for GHL, Sun Ultrasparc
10 (440 MHz) for CGL, Pentium PC 200 MHz for TV, Sun
4/630 MP for WH, Sun Sparc4 IPC for RR, Silicon
Graphics 100 MHz for RT, Silicon Graphics 4D/35 for TA
and Pentium 400 MHz for BB, respectively. Explicit results
for RT and TA have been omitted because no run-time was
provided. It is worth noticing that reported results for WH
includes the best computed solution over five execution runs
as well as related cumulative run-time.
The results shown in Table 1 do not show any conclusive
evidence to support a dominating heuristic over the others.
However, solution quality and run-time reported for BB
proves the HGA-VRP method to be competitive in
comparison to alternative techniques as it mostly matches
the performance of best-known heuristic routing procedures.
Accordingly, the average solution quality deviation (0.48%)
and reasonable run-time obtained certainly shows that
HGAs can be comparable to tabu search techniques.
Evaluating HGA-VRP over a larger statistical sample
involving 10 simulation runs for each data set provided
further evidence of that claim. Results reported in Table 2
indicate average performance, standard deviation and,
relative performance gaps (maximum, average, minimum)
from best-known solutions for each data set, respectively, as
well as overall average deviation. Accordingly, the results
show fairly small overall deviations from best-known

Table 1 Comparison of selected heuristics for VRP

Problem Performance OS GHL CGL TV WH RR BB Best
instance (n) time

1 (50) Dist 524.61 524.61 524.61 524.61 524.61 524.61 524.61 524.61
(min) 1.90 6.0 4.57 0.81 20.0 1.05 2.00
2 (75) Dist 844 835.77 835.45 838.60 835.8 835.32 835.26 835.26
(min) 0.84 53.8 7.27 2.21 50.0 43.38 14.33
3 (100) Dist 838 829.45 829.44 828.56 830.7 827.53 827.39 826.14
(min) 25.72 18.4 11.23 2.39 145.0 36.72 27.90
4 (150) Dist 1044.35 1036.16 1038.44 1033.21 1038.5 1044.35 1036.16 1028.42
(min) 59.33 58.8 18.72 4.51 285.0 48.47 48.98
5 (199) Dist 1334.55 1322.65 1305.87 1318.25 1321.3 1334.55 1324.06 1291.45
(min) 54.10 90.9 28.10 7.50 480.0 77.07 55.41
6 (50) Dist 555.43 555.43 555.43 555.43 555.4 555.43 555.43 555.43
(min) 2.88 13.5 4.61 0.86 30.0 2.38 2.33
7 (75) Dist 911.00 913.23 909.68 920.72 911.8 909.68 909.68 909.68
(min) 17.61 54.6 7.55 2.75 45.0 82.95 10.5
8 (100) Dist 878.00 865.94 866.38 869.48 878.0 866.75 868.32 865.94
(min) 49.99 25.6 11.17 2.90 165.0 18.93 5.05
9 (150) Dist 1184.00 1177.76 1171.81 1173.12 1176.5 1164.12 1169.15 1162.55
(min) 76.26 71.0 19.17 5.67 345.0 29.85 17.88
10 (199) Dist 1441.00 1418.51 1415.40 1435.74 1418.3 1420.84 1418.79 1395.85
(min) 76.02 99.8 29.74 9.11 535.0 42.72 43.86
11 (120) Dist 1043.00 1073.47 1074.13 1042.87 1043.4 1042.11 1043.11 1042.11
(min) 24.07 22.2 14.15 3.18 275.0 11.23 22.43
12 (100) Dist 819.59 819.56 819.56 819.56 819.6 819.56 819.56 819.56
(min) 14.87 16.0 10.99 1.10 95.0 1.57 7.21
13 (120) Dist 1547.00 1573.81 1568.91 1545.51 1548.3 1550.17 1553.12 1541.14
(min) 47.23 59.2 14.53 9.34 510.0 1.95 34.91
14 (100) Dist 866.37 866.37 866.53 866.37 866.4 866.37 866.37 866.37
(min) 19.60 65.7 10.65 1.41 140.0 24.65 4.73

Average deviation from best 1.03% 0.86% 0.69% 0.64% 0.63% 0.55% 0.48%
Average time (min) 33.60 46.8 13.75 3.84 222.85 24.65 21.25
 J Berger and M Barkaoui—Genetic alogorithm for capacitated VRP 1261

Table 2 HGA-VRP average performance

Problem instance (n) Average dist Std dev Max gap Avg gap Min gap

1 (50) 524.61 0 0% 0% 0%
2 (75) 840.07 3.63 1.38% 0.57% 0%
3 (100) 830.027 1.97 0.91% 0.47% 0.15%
4 (150) 1045.21 4.09 2.21% 1.63% 0.91%
5 (199) 1327.16 7.39 3.78% 2.76% 1.87%
6 (50) 555.43 0 0% 0% 0%
7 (75) 909.68 0 0% 0% 0%
8 (100) 868.49 3.10 1.03% 0.29% 0%
9 (150) 1176.44 6.56 1.97% 1.19% 0.25%
10 (199) 1422.71 3.93 2.29% 1.92% 1.45%
11 (120) 1047.54 2.02 0.75% 0.52% 0.09%
12 (100) 819.56 0 0% 0% 0%
13 (120) 1553.95 3.87 1.32% 0.83% 0.38%
14 (100) 866.37 0 0% 0% 0%

Overall average deviation from best-known solutions 1.12% 0.73% 0.36%

Table 3 Comparison of basic and improved HGA-VRP Impact on computational cost was further revealed when
Problem Performance BB BB* Best similar simulations were carried out to assess HGA-VRP*
instance (n) time performance. In that setting, b ran into the interval [1.25, 2]
with Db ¼ 0.25. Results comparing the basic algorithm (BB)
1 (50) Dist 524.61 524.61 524.61 to the improved version (BB*) are reported in Table 3 for the
(min) 2.00 0.20.
2 (75) Dist 835.26 835.83 835.26 same VRP benchmark28 for a single simulation run.
(min) 14.33 13.01 Generally, the use of granular neighbourhoods proves the
3 (100) Dist 827.39 829.44 826.14 algorithm to converge more quickly toward good solutions.
(min) 27.90 15.50 For a near similar quality solution, computational results
4 (150) Dist 1036.16 1037.79 1028.42 show an average run-time reduction on the order of 33%,
(min) 48.98 42.20
5 (199) Dist 1324.06 1321.38 1291.45 representing a significant computational gain. For some
(min) 55.41 35.42 instances, run-time savings can reach up to 50%.
6 (50) Dist 555.43 555.43 555.43
(min) 2.33 1.00
7 (75) Dist 909.68 909.68 909.68
(min) 10.5 7.20 Conclusion
8 (100) Dist 868.32 868.97 865.94
A hybrid genetic algorithm (HGA-VRP) to address the
(min) 5.05 4.35
9 (150) Dist 1169.15 1171.66 1162.55 classical capacitated VRP was presented. Focusing on total
(min) 17.88 7.40 travelled distance minimization, HGA-VRP concurrently
10(199) Dist 1418.79 1420.43 1395.85 evolves two populations of solutions in which respective best
(min) 43.86 44.95 individuals are mutually exchanged through migration over
11(120) Dist 1043.11 1043.18 1042.11
(min) 22.43 10.62 each generation. Genetic operators were designed to
12(100) Dist 819.56 819.56 819.56 incorporate and combine variations of key concepts emer-
(min) 7.21 3.98 ging from recent promising techniques for a time variant of
13(120) Dist 1553.12 1550.56 1541.14 the problem, to further provide search guidance while
(min) 34.91 9.35 maintaining a suitable balance between intensification and
14(100) Dist 866.37 866.37 866.37
(min) 4.73 3.20 diversification. An improved version of the algorithm has
shown to reduce computational cost significantly by further
Average deviation from best 0.48% 0.52% confining exploration of the solution space. Results from a
Average time (min) 21.25 14.17 limited computational experiment showed that HGA-VRP is
cost-effective and very competitive in comparison to the
solutions, 0.73% on average, and around 0.36% (1%) at best-known VRP metaheuristics.
best (at worst). The algorithm also shows some relative Future work will be conducted to improve the proposed
stability when considering the standard deviation average algorithm. Alternative metaheuristic features will be exam-
ratio distribution over data sets (0.7% at worst). Statistical ined to enhance genetic operators while reducing computa-
results from Table 2 provide a new basis for comparative tional cost. Additional improvements might lie on the
and statistical analysis to study future methods. introduction and combination of various selection and
1262 Journal of the Operational Research Society Vol. 54, No. 12
population replacement schemes and, new fitness models.
Application of the approach to related problems will be
explored as well.
References
1 Toth P and Vigo D (2002) (eds). The Vehicle Routing Problem.
SIAM Monographs Discrete Mathematics and Applications.
Society for Industrial and Applied Mathematics, Philadelphia,
USA.
2 Gendreau M, Laporte G and Potvin J-Y (1998). Metaheuristics
for the Vehicle Routing Problem, Les Cahiers du GERAD, G-98–
52. Montreal, Canada.
3 Gendreau M, Laporte G and Potvin J-Y (1997). Vehicle
routing: Modern heuristics. In: Aarts E and Lenstra JK (eds).
Local Search in Combinatorial Optimization. Wiley, Chichester,
pp 311–336.
4 Glover F (1986). Future paths for integer programming and
links to artificial intelligence. Compu Opns Res 13: 533–549.
5 Glover F and Laguna M (1997). Tabu Search. Kluwer
Academic Publishers: Boston.
6 Holland JH (1975). Adaptation in Natural and Artificial Systems.
University of Michigan Press: Ann Arbor.
7 Jong De KA (1975). An analysis of the behavior of a class of
genetic adaptive systems. PhD dissertation, University of
Michigan, USA.
8 Goldberg DE (1989). Genetic Algorithms in Search, Optimiza-
tion, and Machine Learning. Addison-Wesley: New York.
9 Cordeau J-F, Gendreau M, Laporte G, Potvin J-Y and Semet F
(2002). A guide to vehicle routing heuristics. J Opl Res Soc 53:
512–522.
10 Bräysy O and Gendreau M (2004). Vehicle routing problem
with time windows, Part II: metaheuristics to appear in
Transportation Sc. http://www.sintef.no/static/am/opti/pro-
jects/top/.
11 Dalessandro SV, Ochi LS, de A and Drummond LM (1999). A
parallel hybrid evolutionary metaheuristic for the period vehicle
routing problem. In: Rolim JDP et al (eds). Proceedings of the
IPPS/SPDP 1999, 2nd Workshop on Biologically Inspired
Solutions to Parallel Processing Problems, San Juan, Puerto
Rico, Springer, pp 183–191.
12 Berger J, Barkaoui M and Bräysy O (2003). A route-directed
hybrid genetic approach for the vehicle routing problem with
time windows. Inform Systems Opl Res 41: 179–194.
13 Machado P, Tavares J, Pereira F and Costa E (2002). Vehicle
routing problem: doing it the evolutionary way. In: Langdon
WB et al (eds). Proceedings of the Genetic and Evolutionary
Computation Conference New York, USA. Morgan Kaufmann,
San Francisco, p 690.
14 Gehring H and Homberger J (2001). Parallelization of a two-
phase metaheuristic for routing problems with time windows.
Asia-Pacific J Opl Res 18: 35–47.
15 Tan KC, Lee LH and Ou K (2001). Hybrid genetic algorithms
in solving vehicle routing problems with time window con-
straints. Asia-Pacific J Opl Res 18: 121–130.
16 Thangiah SR, Osman IH, Vinayagamoorthy R and Sun T
(1995). Algorithms for the Vehicle Routing Problems with Time
Deadlines. Am J Math Mngt Sci 13: 323–355.
17 Thangiah SR (1995). Vehicle routing with time windows using
genetic algorithms. In: Chambers L. (ed). Application Handbook
of Genetic Algorithms: New Frontiers. Vol. II. CRC Press, Boca
Raton, FL, pp 253–277.
18 Thangiah SR (1995). An adaptive clustering method using a
geometric shape for vehicle routing problems with time
windows. In: Eshelman LJ. (ed). Proceedings of the 6th
International Conference on Genetic Algorithms. Morgan Kauf-
mann, San Francisco, pp 536–543.
19 Blanton JL and Wainwright RL (1993). Multiple vehicle routing
with time and capacity constraints using genetic algorithms. In:
Forrest S. (ed). Proceedings of the 5th International Conference
on Genetic Algorithms. Morgan Kaufmann, San Francisco, pp
452–459.
20 Sangheon Han and Yoshio Tabata (2002). A hybrid genetic
algorithm for the vehicle routing problem with controlling lethal
gene. Int J Asian Pacific Mngt Rev 7: 287–298.
21 Peiris P and Zak SH (2000). Solving vehicle routing problem
using genetic algorithms. Annual Research Summary—Part I—
Research, Section 1—Automatic Control, School of Electrical
and Computer Engineering, Purdue University, http://www.
ece.purdue.edu/ECE/Research/AR S/ARS2000/PART_I/Sec-
tion1/1_19.whtml.
22 Shaw P (1998). Using constraint programming and local search
methods to solve vehicle routing problems. In: Maher M and
Puget J-F (eds.). Principles and Practice of Constraint Program-
ming, Lecture Notes in Computer Science. Springer-Verlag, New
York, pp 417–431.
23 Liu F-H. and Shen S-Y (1999). A route-neighbourhood-based
metaheuristic for vehicle routing problem with time windows.
Eur J Opl Res 118: 485–504.
24 Osman IH (1993). Metastrategy simulated annealing and tabu
search algorithms for the vehicle routing problem. Ann Ops Res
41: 421–451.
25 Toth P and Vigo D (1998). The granular tabu search and its
application to the vehicle routing problem. Technical Report OR/
98/9 DEIS, University of Bologna, Bologna, Italy.
26 Solomon MM (1987). Algorithms for the vehicle routing and
scheduling problems with time window constraints. Ops Res 35:
254–265.
27 Harvey WD and Ginsberg ML (1995). Limited discrepancy
search. In: Mellish CS (ed). Proceedings of the 14th IJCAI
Montreal, Canada. Morgan Kaufmann, San Fransisco, pp 607–
615.
28 Christofides N, Mingozzi A and Toth P (1979). The vehicle
routing problem. In: Christofides N, Mingozzi A, Toth P and
Sandi C (eds). Combinatorial Optimization. Wiley: Chichester,
pp 315–338.
29 Wall M (1995). GAlib—A C þ þ Genetic Algorithms Library,
version 24. http://lancet.mit.edu/galib-2.4/. MIT: Boston.
30 Gendreau M, Hertz A and Laporte G (1994). A tabu search
heuristic for the vehicle routing: problem. Mngt Sci 40: 1276–
1290.
31 Cordeau J-F, Gendreau M and Laporte G (1997). A tabu search
heuristic for the periodic and multi-depot vehicle routing
problems. Networks 30: 105–119.
32 Wark P and Holt J (1994). A repeated matching heuristic for the
vehicle routing problem. J Opl Res Soc 45: 1156–1167.
33 Rego C and Roucairol C (1996). A parallel tabu search
algorithm using ejection chains for the vehicle routing problem.
In: Osman IH and Kelly JP (eds). Meta-Heuristics: Theory and
Applications. Kluwer, Boston, pp 661–675.
34 Rochat Y and Taillard ED (1995). Probabilistic diversification
and intensification in local search for vehicle routing. J
Heuristics 1: 147–167.
35 Taillard ED (1993). Parallel iterative search methods for vehicle
routing problems. Networks 23: 661–673.
Received March 2003;
accepted September 2003 after one rivision