604

Metaheuristics: Heuristic Techniques for

Combinatorial Optimization Problems
Stephan Scheuerer
Fraunhofer Center for Applied Research on Technologies for the Logistics Service Industries ATL, Germany

INTRODUCTION The classical vehicle routing problem (VRP) will be

used for illustration in the following sections. The VRP
Decision support systems (DSSs) provide modern is a hard combinatorial problem with a high relevance
solution techniques that help the decision maker to for the logistics practice. Metaheuristics have proven
find the best solution to a problem. These embedded to successfully tackle this complex problem. In the
solution techniques include and combine, but are not VRP, a set of customers has to be served by a number
limited to, simulation, exact optimization methods, of vehicles located at a depot. Each customer has a
and heuristics. Especially in the field of heuristics, nonnegative demand and a nonnegative service time.
recent advances in metaheuristic methods have proved The VRP consists of designing a set of cost-minimal
to be remarkably effective so that metaheuristics are vehicle routes, each starting and ending at the depot.
nowadays the preferred way for solving many types of Each customer must be visited exactly once by a vehicle,
complex problems, particularly those of combinatorial with the total demand of any route not exceeding the
nature. Some of these problems are, for example, the vehicle capacity, and the total duration of any route not
well-known “traveling salesman” problem, the gener- surpassing a preset limit. For a comprehensive overview
alized assignment problem, the set-covering problem, on the VRP, refer to Toth and Vigo (2002).
and vehicle and network routing applications. Most of
all, metaheuristics allow us to solve real-world prob-
lems with a notably high level of complexity. This is BACKGROUND
where exact methods are often incapable of finding
solutions whose qualities are close to that obtained by The application of exact optimization methods to
the leading metaheuristics. Metaheuristic applications real-life problems assumes that it is possible to find
with world-class performance can be found in all kinds an optimal solution in a reasonable computing time.
of areas such as economics, engineering, and natural This, however, is not the case for most real-life combi-
sciences. natorial optimization problems1 like, for example, rich
This article will describe the basic notions of several vehicle routing or scheduling problems. To overcome
metaheuristic techniques to show the user of a DSS this shortcoming, heuristic methods were developed
how the principles of the embedded solution techniques to help the decision maker find a satisfying solution,
could apply. For the sake of brevity, the following dis- in the best case the optimum, in a realistic amount of
cussion will concentrate on a representative sample of computing time.
the different heuristic paradigms, namely, tabu search In the early years, specialized heuristics were de-
(TS), simulated annealing (SA), greedy randomized veloped for a range of specific optimization problems.
adaptive search procedure (GRASP), ant colony op- In this context, the knowledge about the optimization
timization (ACO), evolutionary algorithms (EA), and problem itself plays a significant role in the design
scatter search (SS). For a more detailed insight into of the heuristic, especially for the solution quality,
this emerging research area, the following small subset as most heuristic procedures in this category have a
out of the overwhelming number of literature on this constructive nature. These constructive methods build
topic could provide a starting point for further reading: up a complete solution from scratch by sequentially
Gendreau and Potvin (2005), Glover and Kochenberger adding components to a solution fragment until the
(2003), and Blum and Roli (2003). solution is complete. This sequential process often

Copyright © 2008, IGI Global, distributing in print or electronic forms without written permission of IGI Global is prohibited.
Metaheuristics: Heuristic Techniques for Combinatorial Optimization Problems

leads to good, rarely to optimal, solutions. To improve than a single solution. Single-solution metaheuristics
the solution, so-called improvement methods can be may also be called trajectory methods because the M
applied. A technique that is used very often in the class search is characterized by a trajectory in the search
of improvement methods is local search algorithms space (Blum & Roli, 2003).
that try to iteratively modify a given solution in order In metaheuristic search, the goal is an efficient and
to generate improvements. Modifications are done on effective exploration of the search space. In order to
a local basis, for example, by shifting or exchanging a achieve this aim, the right balance between intensifica-
single component in a complete solution. These kinds tion and diversification is crucial. Intensification refers
of local operations (often called moves) could lead to to a more thorough examination of attractive-looking
a new best solution.2 All solutions that can be reached regions whilst diversification encourages the search for
by a given kind of modification are considered to build unvisited regions of the search space and should thus
up the set of neighboring solutions, called the neigh- lead to new and even more attractive regions. A well-
borhood, to the given solution. A descent local search balanced use of these two elements can be seen as the
algorithm performs these modifications as long as there key to the design of efficient metaheuristics.
is an improvement in the objective function value. The
search stops whenever there is no further improve-
ment. The solution is then considered to be a local OVERVIEW OF METAHEURISTIC
optimum with respect to the neighborhood considered. PRINCIPLES
An example for a constructive method in the context of
vehicle routing is the well-known savings algorithm of In the following, some metaheuristics will be briefly
Clarke and Wright (1964). Starting from single round described. The reader is referred to the literature at the
trips from the depot to each customer, the routes are end of each section for further readings.
merged according to a savings criterion that expresses
the distance saved through this step as long as feasibility Tabu Search (TS)
of the solution is maintained. To improve the solution
even more (sometimes still during the construction TS was first introduced by Glover (1986) and Hansen
phase), for example, k-opt improvement routines3 can (1986). A simple TS can be seen as a local search strat-
be applied. This method iteratively modifies up to k egy where always the best solution in the neighborhood
edges of a single vehicle tour as long as there is an of the current solution is selected as the new solution,
improvement in tour cost. even if it deteriorates solution cost. This strategy helps
Using local search methods often helps to make to overcome local optima. To prevent the search from
significant improvements to an initially constructed cycling, recently visited solutions are considered for-
solution. However, they have the disadvantage of pos- bidden or tabu for a certain number of iterations. A
sibly being trapped in a local optimum that is not the short-term memory, called a tabu list, is used to store
global optimum. To solve this problem, metaheuristic these solutions and characteristic attributes of these
techniques can be applied. The way of achieving this solutions. Aspiration criteria define rules that allow us
aim depends on the philosophy of the metaheuristic. to overcome the tabu status under certain conditions.
The term metaheuristic was originally coined by The algorithm stops when a termination condition, for
Glover (1986) and describes a more general, not prob- example, a maximum number of iterations, is met.
lem-specific solution scheme. Metaheuristics can be For the VRP, a move can be a shift of a single cus-
considered4 strategies that guide a search process to tomer into another vehicle’s tour. The reverse move back
efficiently explore the search space for finding a (close into its original tour would be forbidden for a number
to) optimal solution. of iterations (tabu tenure). The reverse move would,
One possibility5 for classifying metaheuristics is however, be allowed if it led to a new best solution
dividing them into the following two categories: single- (aspiration criteria). A whole number of successful TS
solution metaheuristics (representatives are, e.g., TS, implementations for the VRP can be reported.6
SA, GRASP), which basically work on a single solution One crucial issue in the design of tabu search algo-
at a time, and population-based metaheuristics (e.g., rithms is the tabu tenure. Too small tabu tenures cannot
EA, SS, ACO), which consider a set of solutions rather effectively prevent the search from cycling. However,

605
 Metaheuristics: Heuristic Techniques for Combinatorial Optimization Problems

large tabu tenures make the search visit new regions of the initial solution is generated by means of the sav-
the search space, but may prevent a good examination ings algorithm of Clarke and Wright (1964). He uses
of an attractive region. To balance and to strategically an advanced neighborhood structure, where up to λ
use these opposite tendencies, different tabu tenure customers can be exchanged between two different
strategies were designed. Other mechanisms that try routes, and proposes a specific cooling schedule for
to overcome this problem and that aim at finding the the SA algorithm.
right balance of intensification and diversification
are, for example, the restriction of the neighborhood Greedy Randomized Adaptive Search
by building so-called candidate lists or the usage of Procedure (GRASP)
long-term memory that records and penalizes solution
attributes that have often been found in previously The basic idea of GRASP consists of finding a global
visited solutions. optimal solution by repeatedly applying a local search
Introductions and descriptions of advanced TS to diverse starting solutions, each constructed by a fast
topics may be found in Glover and Laguna (1997) and greedy construction method. This strategy is commonly
Gendreau (2002, 2003). used by so-called multistart methods. The multistart
method GRASP, however, is characterized by a ran-
Simulated Annealing (SA) domized greedy construction process, where, at each
step of the construction process, the element to be
SA, first proposed by Kirkpatrick, Gelatt, and Vecchi included in the solution fragment is chosen randomly
(1983) and Cerny (1985), is a local search procedure from a list (called the restricted candidate list, RCL)
that randomly chooses a solution out of the neighbor- of the most promising elements not yet incorporated
hood of the current solution at each iteration. The into the partial solution. The random selection triggers
solution is taken as a new solution if it either improves the construction of different solutions.
the solution value or if it is accepted due to an accep- A number of strategies for the selection of ele-
tance criterion. The acceptance criterion, called the ments and the size of the RCL have been suggested.
metropolis criterion, is probabilistic and relative to the Some recent developments incorporate some form of
magnitude of the cost increase and to the temperature memory to overcome the drawback that each restart is
parameter. The smaller the cost deterioration and the independent of previous ones or attempts to bias the
higher the temperature, the higher the probability of probability distribution associated with the elements
acceptance. The temperature is decreased during the in the RCL. Details on GRASP techniques and ap-
search process according to a cooling scheme. The plications may be found in Festa and Resende (2002),
algorithm was inspired by the physical annealing pro- Pitsoulis and Resende (2002), as well as in Resende
cess in condensed-matter physics, where first, a solid and Ribeiro (2003).
is melted and then, the temperature is progressively
reduced to recover a solid state of lower energy. Further Ant Colony Optimization (ACO)
details on this method can be found in Aarts, Korst,
and van Laarhoven (1997) and Henderson, Jacobson, ACO, first proposed by Dorigo (1992) in a basic form
and Johnson (2003). called Ant System, is a metaheuristic that is inspired
The choice of the appropriate cooling scheme is one by the behavior of real ants. To find its path from food
of the crucial issues in SA algorithms. Temperature has sources to the nest and vice versa, each ant marks its
to be reduced carefully and through a series of levels. trail with a substance called pheromone. The stronger
Different forms of static and dynamic cooling schemes the pheromone concentration on a trail, the higher the
as well as more general forms of acceptance rules have probability that ants will follow this path. This coop-
been suggested. One well-known deterministic vari- erative behavior of the ants leads to the emergence
ant is threshold accepting (Dueck & Scheuer, 1990). of shortest paths. In ACO, artificial ants sequentially
This method determines the acceptance criterion by construct solutions in a greedy fashion with a random
a predefined upper-bound value that is progressively selection of the next element to be incorporated into
reduced. The method has been applied to the VRP, for the partial solution. Here, the selection criterion is in-
example, by Osman (1993). In this implementation, fluenced by the amount of pheromone associated with

606
Metaheuristics: Heuristic Techniques for Combinatorial Optimization Problems

each element. An element with a good heuristic evalu- basic idea is the iterative selection of solutions out
ation and a high degree of pheromone is more likely to of the current population (called parents) to generate M
be selected. ACO’s basic form describes that as soon new solutions (called offspring), and then to update the
as all ants have constructed a solution, the pheromone population with these new solutions. The most widely
levels of the elements are updated by increasing the known representatives of this type of metaheuristics are
level for solution components that have been found genetic algorithms (GAs; Holland, 1975). In basic GAs,
in high-quality solutions. The search is then restarted the selection process of the parents is probabilistically
with the updated pheromone levels until a stopping biased toward the best solutions in the population. New
criterion is met. solutions are generated by the application of operators
For the VRP, for example, Reimann, Doerner, and similar to those found in natural selection, namely,
Hartl (2004) developed an ant system on the basis crossover and mutation. The crossover operator tries
of the well-known savings algorithm of Clarke and to combine the most desirable features of two parent
Wright (1964). In a first phase, the savings-based pro- solutions to create attractive offspring solutions. The
cedure is used to generate a pool of good solutions. mutation operator may perform a small random pertur-
An improvement mechanism is then applied to each of bation to the offspring before the current population is
these solutions. Instead of using the classical savings updated with the new offspring solutions. This process
measure in the savings algorithm, the authors work is repeated until a stopping criterion is met. The best
with an attractiveness value. This value additionally solution found is returned.
contains information about how good a combination In their original forms, GAs encode solutions as
of two elements turned out to be in previous iterations. bit or integer strings, referred to as chromosomes. The
This information is provided by pheromone levels for choice of an appropriate representation is crucial for
a pair of elements. After all ants have constructed their the success of an EA. As the encoding itself may be a
solutions, the pheromone trails are updated on the basis complicated task, a range of implementations ignore
of the solutions found by the ants. In a second phase the typical encoding. Modern implementations also
of the algorithm, the best solution is decomposed into apply specialized operators, for example, including
several subproblems, each of which is optimized by local search operators as a special form of the muta-
means of the procedure in the first phase. The algorithm tion operator to improve the offspring solution and,
repeatedly applies the two phases until a stopping in case of infeasibility, an operator that tries to repair
criterion is reached. the solution.
Different pheromone update schemes have been sug- Prins (2004), for example, proposes a GA including
gested. For example, to avoid a too rapid convergence of a local search procedure for the VRP. In his approach, a
the algorithm in a suboptimal region of the search space chromosome is a sequence (permutation) of customers,
and to incorporate a form of diversification, pheromone without any trip delimiters. It may be interpreted as
concentrations may be periodically reduced to simulate the order in which a vehicle must visit all customers if
a form of evaporation. To intensify the search in an at- the same vehicle performs all trips one by one without
tractive region, the amount of pheromone for certain going through the depot. At any time, a chromosome
elements may also be increased by some external factor. can be converted into an optimal VRP solution (subject
Recent developments in ACO include approaches that to chromosome sequence) due to a special splitting
bound the pheromone levels as well as the coupling of procedure. This design choice avoids repair procedures
ACO with local search algorithms. For more details on and enables the use of classical crossover operators. In
developments in ACO, see Dorigo and Stützle (2003, contrast, GA implementations based on chromosomes
2004) as well as Maniezzo and Carbonaro (2002). with trip delimiters need a repair procedure to get fea-
sible children after each crossover.
Evolutionary Algorithms (EAs) Due to its structure, EAs as well as all population-
based methods tend quite naturally to parallel implemen-
EAs are based on the Darwinian principle of natural tations. One popular way7 is distributing subpopulations
selection. They work with a set of solutions, called of solutions to different processors and ensuring some
the population, and are initially created randomly or form of synchronization between them.
heuristically rather than with a single solution. The

607
 Metaheuristics: Heuristic Techniques for Combinatorial Optimization Problems

EAs have been applied to most combinatorial optimi- Path relinking (PR) is closely related to SS. It is a
zation problems with great success. An overwhelming strategy that generates new solutions through build-
number of literature on this metaheuristic exists; see, ing trajectories between a starting solution and one or
for example, Reeves (2003), Hertz and Kobler (2000), more target solutions that guide a search process. See
Mitchell (1998), and Michalewicz (1996). Laguna and Martí (2003), Glover, Laguna and Martí
(2000, 2003), and Glover (1998) for more details on
Scatter Search (SS) PR and SS.

In its basic form, SS is closely related to EA. Glover

(1998) and Laguna and Martí (2003) present a general FUTURE TRENDS
template for SS. In most SS implementations, the set of
solutions, called the reference set (called the population Over the years, more and more metaheuristic imple-
in EAs), is divided into a subset of high-quality solutions mentations integrate elements that were originally
and a subset of diverse solutions. The reference set is proposed in conjunction with other methods. This has
initially filled with a diverse set of heuristically gener- led to hybrid heuristics,8 which use elements of different
ated solutions. The main difference between SS and methods to combine their strengths. This effect may be
EA is the fact that SS provides unifying principles for seen as a natural development of metaheuristic design
the selection and combination of solutions. That means based on the fact that all metaheuristic strategies have
that in SS, a variable number of parent solutions may some specific strengths and weaknesses.
be used to create new solutions (whereas in classical The next step in this development process may be a
GA, only two parents are considered). Furthermore, common view on metaheuristics that is now emerging.
the combination of parent solutions is not limited to Some research has be done in this area, for example, by
the classical EA operators but instead allows more Taillard, Gambardella, Gendreau, and Potvin (2001),
flexible forms of combinations. The underlying idea who provide an algorithmic framework called adaptive
is to build linear combinations of solutions that lead memory programming. Blum and Roli (2003) make a
to new solutions inside and outside the convex region conceptual comparison of metaheuristics based on the
spanned by the reference solutions. The generated solu- principles of intensification and diversification. Gen-
tions, called trial solutions, are improved or repaired, dreau and Potvin (2005) identify a unifying framework
and the best ones, according to quality and diversity, with a number of common algorithmic components
are considered for an update of the reference set. The found in metaheuristics.
search stops whenever a stopping criterion is met or The trend is (or at least should be) also toward
no more new solutions can be generated. For the lat- simpler and more flexible heuristics that are still highly
ter, it is possible, for example, to restart the algorithm efficient and effective, but have fewer parameters, and
by replacing solutions (except the best ones) from the which are easily adaptable to new practical problems
reference set with new diverse solutions. and further side constraints.
The scatter search heuristic is, for example, used
by Corberan, Fernandez, Laguna, and Martí (2000) to
address the problem of routing school buses consider- CONCLUSION
ing multiple objectives. They developed two construc-
tion heuristics to initially build several solution pools A lot of complex decisions in areas such as econom-
dependant on the number of routes in the solution. ics, engineering, or natural sciences can be described
The scatter search principle is applied to each of these as combinatorial optimization problems. Finding a
pools using a combination method that first matches solution to these problems is quite a challenge. Over
the routes from one solution to the routes of another the years, heuristic techniques named metaheuristics
solution. The rationale behind this is to find the com- have emerged. They allow us to efficiently tackle these
mon structure of the two solutions being combined to problems. This article is an attempt to provide DSS
transfer this structure to the newly generated solution. engineers as well as decision makers with the basic
After matching, a newly combined solution is generated principles of these powerful algorithmic approaches.
by a voting mechanism.

608
Metaheuristics: Heuristic Techniques for Combinatorial Optimization Problems

Gendreau, M. (2002). Recent advances in tabu search.

REFERENCES In C. C. Ribeiro & P. Hansen (Eds.), Essays and surveys M
in metaheuristics (pp. 369-378). Boston: Kluwer.
Aarts, E. H. L., Korst, J. H. M., & van Laarhoven, P.
Gendreau, M. (2003). An introduction to tabu search.
J. M. (1997). Simulated annealing. In E. H. L. Aarts
In F. Glover & G. A. Kochenberger (Eds.), Handbook
& J. K. Lenstra (Eds.), Local search in combinatorial
of metaheuristics (pp. 37-54). Boston: Kluwer.
optimization (pp. 91-136). Chichester, United King-
dom: Wiley. Gendreau, M., Laporte, G., & Potvin, J.-Y. (2002).
Metaheuristics for the capacitated VRP. In P. Toth &
Blum, C., & Roli, A. (2003). Metaheuristics in com-
D. Vigo (Eds.), The vehicle routing problem (pp. 129-
binatorial optimization. ACM Computing Surveys,
154). Philadelphia: SIAM Monographs on Discrete
35(3), 286-308.
Mathematics and Applications.
Cerny, V. (1985). Thermodynamical approach to the
Gendreau, M., & Potvin, J.-Y. (2005). Metaheuristics
traveling salesman problem: An efficient simulation
in combinatorial optimization. Annals of Operations
algorithm. Journal of Optimization Theory and Ap-
Research, 140, 189-213.
plications, 45, 41-45.
Glover, F. (1986). Future paths for integer program-
Clarke, G., & Wright, J. W. (1964). Scheduling of
ming and links to artificial intelligence. Computers &
vehicles from a central depot to a number of delivery
Operations Research, 13, 533-549.
points. Operations Research, 12, 568-581.
Glover, F. (1998). A template for scatter search and
Corberan, A., Fernandez, E., Laguna, M., & Martí, R.
path relinking. In J. K. Hao, E. Lutton, R. Ronald, M.
(2000). Heuristic solutions to the problem of routing
Schoenauer, & D. Snyers (Eds.), Artificial evolution
school buses with multiple objectives. Journal of the
(LNCS 1363, pp. 13-54). Berlin, Germany: Springer.
Operational Research Society, 53, 427-435.
Glover, F., & Kochenberger, G. A. (Eds.). (2003).
Crainic, T. G., & Toulouse, M. (2003). Parallel strategies
Handbook of metaheuristics. Boston: Kluwer.
for meta-heuristics. In F. Glover & G. A. Kochenberger
(Eds.), Handbook of metaheuristics (pp. 475-513). Glover, F., & Laguna, M. (1997). Tabu search. Boston:
Boston: Kluwer. Kluwer.
Dorigo, M. (1992). Optimization, learning, and natural Glover, F., Laguna, M., & Marti, R. (2000). Funda-
algorithms. Unpublished doctoral dissertation, Politec- mentals of scatter search and path relinking. Control
nico di Milano, Milan. and Cybernetics, 39, 653-684.
Dorigo, M., & Stützle, T. (2003). The ant colony opti- Glover, F., Laguna, M., & Marti, R. (2003). Scatter
mization metaheuristic: Algorithms, applications, and search and path relinking: Advances and applications.
advances. In F. Glover & G. A. Kochenberger (Eds.), In F. Glover & G. A. Kochenberger (Eds.), Handbook
Handbook of metaheuristics (pp. 251-285). Boston: of metaheuristics (pp. 1-35). Boston: Kluwer.
Kluwer.
Hansen, P. (1986). The steepest ascent mildest descent
Dorigo, M., & Stützle, T. (2004). Ant colony optimiza- heuristic for combinatorial programming. Paper
tion. Cambridge, MA: MIT Press. presented at the Congress on Numerical Methods in
Combinatorial Optimization, Capri, Italy.
Dueck, G., & Scheuer, T. (1990). Threshold accepting:
A general purpose optimization algorithm. Journal of Henderson, D., Jacobson, S. H., & Johnson, A. W.
Computational Physics, 90, 161-175. (2003). The theory and practice of simulated annealing.
In F. Glover & G. A. Kochenberger (Eds.), Handbook
Festa, P., & Resende, M. C. G. (2002). GRASP: An
of metaheuristics (pp. 287-319). Boston: Kluwer.
annotated bibliography. In C. C. Ribeiro & P. Hansen
(Eds.), Essays and surveys in metaheuristics (pp. 325- Hertz, A., & Kobler, D. (2000). A framework for the
368). Boston: Kluwer. description of evolutionary algorithms. European
Journal of Operational Research, 126, 1-12.

609
 Metaheuristics: Heuristic Techniques for Combinatorial Optimization Problems

Holland, J. H. (1975). Adaption in natural and artificial Resende, M. G. C., & Ribeiro, C. C. (2003). Greedy
systems. Ann Harbor, MI: The University of Michigan randomized adaptive search procedures. In F. Glover &
Press. G. A. Kochenberger (Eds.), Handbook of metaheuristics
(pp. 219–249). Boston: Kluwer.
Kirkpatrick, S., Gelatt, C. D., & Vecchi, M. P. (1983).
Optimization by simulated annealing. Science, 220, Taillard, E. D., Gambardella, L.-M., Gendreau, M., &
671-680. Potvin, J.-Y. (2001). Adaptive memory programming:
A unified view of metaheuristics. European Journal of
Laguna, M., & Martí, R. (2003). Scatter search: Meth-
Operational Research, 135, 1-16.
odology and implementations in C. Boston: Kluwer.
Talbi, E. G. (2002). A taxonomy of hybrid metaheuris-
Laporte, G., & Semet, F. (2002). Classical heuristics
tics. Journal of Heuristics, 8, 541-564.
for the VRP. In P. Toth & D. Vigo (Eds.), The vehicle
routing problem (pp. 109-128). Philadelphia: SIAM Toth, P., & Vigo, D. (Eds.). (2002). The vehicle routing
Monographs on Discrete Mathematics and Applica- problem. Philadelphia: SIAM Monographs on Discrete
tions. Mathematics and Applications.
Lawler, E. L. (1976). Combinatorial optimization:
Networks and matroids. New York: Holt, Reinhart
and Winston. KEY TERMS
Maniezzo, V., & Carbonaro, A. (2002). Ant colony
Constructive Methods: Constructive methods
optimization: An overview. In C. C. Ribeiro & P.
are heuristics that build up a complete solution from
Hansen (Eds.), Essays and surveys in metaheuristics
scratch by sequentially adding components to a partial
(pp. 469-492). Boston: Kluwer.
solution until the solution is complete.
Michalewicz, Z. (1996). Genetic algorithms + data
structures = evolution programs (3rd ed.). Berlin, Ger- Diversification: Diversification is a strategy that
many: Springer. encourages a search process to examine unvisited
regions of the search space.
Mitchell, M. (1998). An introduction to genetic algo-
rithms. Cambridge, MA: MIT Press.
Improvement Methods: These are methods that
Osman, I. H. (1993). Metastrategy simulated annealing attempt to improve a given solution either in cost or
and tabu search algorithms for the vehicle routing prob- in structure.
lem. Annals of Operations Research, 41, 421-451.
Intensification: Intensification is a strategy that
Pitsoulis, L. S., & Resende, M. G. C. (2002). Greedy encourages a thorough examination of attractive-look-
randomized adaptive search procedure. In P. Pardalos & ing regions in the search space.
M. Resende (Eds.), Handbook of applied optimization
(pp. 168-183). New York: Oxford University Press. Metaheuristics: These are strategies that guide a
Prins, C. (2004). A simple and effective evolutionary heuristic search process to efficiently explore the search
algorithm for the vehicle routing problem. Computers space in order to find a (close to) optimal solution.
& Operations Research, 31, 1985-2002.
Population-Based Metaheuristics: These are
Reeves, C. R. (2003). Genetic algorithms. In F. Glover & metaheuristics that consider a set of solutions rather
G. A. Kochenberger (Eds.), Handbook of metaheuristics than a single solution.
(pp. 55-82). Boston: Kluwer.
Reimann, M., Doerner, K., & Hartl, R. F. (2004). D- Single-Solution Metaheuristics: These are me-
Ants: Savings based ants divide and conquer the vehicle taheuristics that basically work on a single solution
routing problem. Computers & Operations Research, at a time.
31, 563-591.

610
Metaheuristics: Heuristic Techniques for Combinatorial Optimization Problems

ENDNOTES 3
See Laporte and Semet (2002) for explanations
on constructive and improvement heuristics for M
1
Combinatorial optimization can be characterized the VRP.
as the mathematical attempt of determining the 4
See Blum and Roli (2003) for an overview on
optimal sequencing, grouping, or selection of possible definitions.
a discrete, limited number of objects (Lawler, 5
See Blum and Roli (2003) for an overview on
1976). possible definitions.
2
Note that in the article, we refer to a minimiza- 6
See Gendreau, Laporte, and Potvin (2002) for an
tion problem to simplify the description. For a overview on metaheuristics for the VRP.
minimization problem, a solution is considered 7
See Crainic and Toulouse (2003) for parallel
to be better than another if it decreases solution strategies for metaheuristics.
cost. 8
See Talbi (2002) for a taxonomy of hybrid heu-
ristics.

611