See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/223369558

The pickup and delivery problem with time windows

Article in European Journal of Operational Research · September 1991

DOI: 10.1016/0377-2217(91)90319-Q

CITATIONS READS

669 8,326

3 authors, including:

Jacques Desrosiers François Soumis

HEC Montréal Polytechnique Montréal
131 PUBLICATIONS 13,613 CITATIONS 179 PUBLICATIONS 10,615 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Jacques Desrosiers on 05 January 2018.

The user has requested enhancement of the downloaded file.

 A Grouping Genetic Algorithm
for the Pickup and Delivery Problem
with Time Windows
Giselher Pankratz
Diskussionsbeitrag Nr. 337
Mai 2003

Diskussionsbeiträge des Fachbereichs Wirtschaftswissenschaft

der FernUniversität in Hagen
Herausgegeben vom Dekan des Fachbereichs
 A Grouping Genetic Algorithm for the Pickup and Delivery Problem with Time Windows

Giselher Pankratz

Abstract:
The Pickup and Delivery Problem with Time Windows (PDPTW) is a generalization of the
well studied Vehicle Routing Problem with Time Windows (VRPTW). Since it models sev-
eral typical planning situations in operational transportation logistics and public transit, the
PDPTW has attracted growing interest in recent years. This paper proposes a Grouping Ge-
netic Algorithm for solving the PDPTW which features a group-oriented genetic encoding in
which each gene represents a group of requests instead of a single request. The GGA is sub-
ject to a comparative test on the basis of two publicly available benchmark problem sets that
comprise 9 and 56 PDPTW instances, respectively. The results show that the GGA is able to
consistently find high quality solutions.

Fachbereich Wirtschaftswissenschaft, FernUniversität in Hagen

Profilstr. 8, D-58084 Hagen, BRD

Tel.: 02331/987-4432
Fax: 02331/987-4447
E-Mail: [email protected]
 1

A Grouping Genetic Algorithm

for the Pickup and Delivery Problem with Time Windows

Giselher Pankratz

1 Introduction
The Pickup and Delivery Problem with Time Windows (PDPTW) can be described as follows
(Savelsbergh and Sol, 1995): A set of transportation requests that is known in advance has to
be satisfied by a given fleet of vehicles. Each request is characterized by its pickup location
(origin), its delivery location (destination) and the size of the load that has to be transported
from the origin to the destination. For each pickup and delivery location, a time window and
loading and unloading times are specified. The load capacity, the maximum length of its
operating interval, a start location and an end location are given for each vehicle. In order to
fulfill the requests, a set of routes has to be planned such that each request is transported from
its origin to its destination by exactly one vehicle. Basically, the PDPTW differs from the
VRPTW by the additional precedence constraints, i.e. the restriction that the origin of each
request has to be visited before the corresponding destination.
In accordance with definitions of the PDPTW usually found in the literature and in order
to make the proposed approach comparable to other recently published approaches, it is as-
sumed that the number of vehicles is unlimited and all vehicles have the same capacity and the
same home location where all routes start and end. A reasonable objective function may use
optimization criteria such as, number of vehicles employed, total distance traveled, total sched-
ule duration or combinations of these.
Formulations of the PDPTW as an integer program have been presented by, e.g., Dumas
et al. (1991) and Savelsbergh and Sol (1995). Because this contribution concentrates on solu-
tion approaches, a review of formal models for the PDPTW is omitted here. Since the PDPTW
is a generalization of the Vehicle Routing Problem with Time Windows (VRPTW), it is at least
as complex as the latter, which has been proven by Lenstra and Rinnoy Kan (1981) to be NP-
hard.
The PDPTW models a variety of operational planning problems in transportation logis-
tics. Applications range from local area courier services to less-than-truckload transportation
and long-distance haulage. In addition, the PDPTW also matches typical situations in public
transit. For example, the Dial-a-Ride Problem with Time Windows (DARPTW) is a PDPTW in
which people instead of goods are to be transported. While in the past research mainly concen-
trated on the VRPTW, increasing interest in the PDPTW can be observed in recent years.
Nevertheless there is still relatively little attention paid to the PDPTW in the literature.
The paper is organized as follows. Section 2 gives a review of previous work on the
PDPTW and similar problems. In Section 3 the proposed solution approach, a Grouping Ge-
 2

netic Algorithm (GGA), is described in detail. Section 4 reports computational results obtained
by the GGA for two publicly available benchmark data sets. Finally, the main findings are
summarized in section 5.

2 Literature Review
In this section, an overview of important previous approaches for solving the PDPTW is given.
Compared to the comprehensive review of models and approaches of Savelsbergh and Sol
(1995), a rather brief survey is presented here, taking into account some more recent develop-
ments. The methods can be divided into optimal approaches, problem specific heuristics and
metaheuristics.

2.1 Optimal approaches

The first optimization algorithms for the PDPTW were restricted to the single vehicle case,
which is considerably easier to solve than the regular PDPTW with multiple vehicles. Impor-
tant approaches of this kind are the dynamic programming algorithms by Psaraftis (1980,
1983a) and Desrosiers et al. (1986). In the 1990s, the first exact approaches for the multi-
vehicle PDPTW were published. Based on a formulation as a set partitioning problem, Dumas
et al. (1991) proposed a branch-and-price algorithm for the multiple-vehicle PDPTW. Feasible
routes are found by solving a constrained shortest path problem. As the authors point out, the
algorithm is appropriate for instances in which each request occupies a relatively large per-
centage of vehicle capacity. By means of this algorithm, they were able to optimally solve two
practical problems with 19 and 30 requests respectively, which at 40 per cent, represent full
truck loads. Savelsbergh and Sol (1998) also developed a branch-and-price algorithm for the
multiple-vehicle PDPTW. In this algorithm, several sophisticated techniques, such as embed-
ded heuristics and a special column management scheme, are employed in order to speed up
search. Optimal solutions to several test cases with 30 requests each are reported. Since Savels-
bergh and Sol observe significant differences in solution times, even for instances in the same
problem class, they classify their optimization algorithm as unsuitable for practical planning
situations.

2.2 Problem specific approximation methods

Due to the complexity of the PDPTW, several heuristic algorithms for the PDPTW have been
published. Some of them construct solutions following a “cluster first, route second” approach.
For example, Bodin and Sexton (1986) propose an iterative procedure, which, in each iteration,
first (re)assigns requests to vehicles and then solves a single-vehicle PDPTW for each vehicle.
A more fine-grained approach is presented by Desrosiers et al. (1988) who introduce the notion
 3

of a mini-cluster, i.e. a feasible route segment that serves one or more requests starting and
ending with an empty vehicle. Each mini-cluster is treated as a single full truck load. After a
large number of mini-clusters has been generated heuristically, a set partitioning problem is
solved. Ioachim et al. (1995) pick up the idea of mini-clusters and present a method that
employs an exact algorithm to produce an optimal set of mini-clusters.
Since insertion heuristics have proven successful for the VRPTW (Solomon, 1987), it ap-
pears reasonable to apply similar principles to the PDPTW as well. A frequently cited insertion
procedure for the DARPTW, which can be easily adapted to the PDPTW, is presented by Jaw
et al. (1986). Madsen et al. (1995) develop an insertion algorithm for the DARPTW which is a
generalized and improved version of the heuristic of Jaw et al.
In order to improve the quality of solutions obtained by a construction heuristic, Van der
Bruggen et al. (1993) propose a local improvement procedure for the single-vehicle PDPTW
based on arc exchanges following the variable-depth search procedure of Lin and Kernighan
(1973) for the Traveling Salesman Problem (TSP). Similar arc exchange operators for routing
problems with precedence constraints are described by Psaraftis (1983b). In contrast to these
arc-centered approaches, Toth and Vigo (1996) choose a request-oriented improvement
method. They distinguish between intra-route movements that only affect the relative order of
the locations in a route and inter-route movements that shift requests between routes.

2.3 Metaheuristics
Problem specific heuristics hold the danger of the search being trapped in a local optimum.
Metaheuristics have become widely accepted as an appropriate means of avoiding this undesir-
able behavior. Compared to the VRPTW for which a considerably large number of metaheu-
ristic approaches is reported in the literature, until now only a few metaheuristics have been
reported for the PDPTW. Most of them are adaptations of the tabu search algorithm.
In the tabu search heuristic of Gendreau et al. (1998) the neighborhood definition is based
on the concept of ejection chains. In an ejection chain, a request is eliminated from one route
and inserted into another route, which in turn causes a request from that route to be shifted to a
third route, and so on. As a diversification strategy, the authors make use of a so-called adap-
tive memory, which contains a number of the best solutions found thus far. After the search has
terminated, a new solution is constructed from the routes in the memory, serving as a new
starting solution.
Based on the definition of three distinct moves, Nanry and Barnes (2000) propose a reac-
tive tabu search for the PDPTW in which the search progress is continuously analyzed and
search parameters such as the length of the tabu list are adjusted accordingly during search.
Additionally, a special procedure is employed in order to detect and escape from chaotic attrac-
tor basins in the search space. In order to evaluate their approach, the authors use problem in-
stances that were derived from Solomon test cases for the VRPTW (Solomon, 1987) for which
 4

optimal solutions are known. During problem transformation, for each PDPTW instance, the
feasibility of its corresponding optimal VRPTW solution was maintained. Thus, the results can
be directly compared to the optimal VRPTW solutions. It should be noted that Nanry and Bar-
nes were the first to make their PDPTW benchmark problems publicly available.
Yet another tabu search is presented by Lau and Liang (2001). A special construction al-
gorithm, called a “partitioned insertion heuristic”, is used in order to build starting solutions.
The neighborhood definitions in this approach are adapted from Nanry and Barnes (2000). In-
stead of moving or exchanging single requests, they allow the exchange of so-called clusters,
which are very similar to the notion of mini-clusters introduced by Desrosiers et al. (1988). The
approach is tested on the nine 100-node instances provided by Nanry and Barnes (2000). Fur-
thermore, the authors give a detailed description of a transformation scheme by which they
generate additional PDPTW instances based on a selection of another 27 Solomon VRPTW
instances. The results for these instances are compared to the best, near-optimal VRPTW solu-
tions that have been published.
Li and Lim (2001) developed a hybrid metaheuristic called tabu-embedded simulated an-
nealing. Three different request swapping moves are used to define local search neighborhood
structures. Basically, a simulated annealing procedure is performed, and is restarted from the
current best solution every time a given number of iterations without improvement has been
observed. In addition, a tabu list keeps track of the recently visited solutions in order to avoid
cycling. The approach is evaluated using the nine 100-node test cases of Nanry and Barnes.
Additionally, Li and Lim report results on 56 own problem instances generated on the basis of
the full set of all 56 Solomon VRPTW instances with 100 customers and the respective best
solutions from the literature.
Despite the success of Genetic Algorithms for the VRPTW (see e.g. Blanton and Wain-
wright, 1993; Thangiah, 1995; Potvin and Bengio, 1996; Berger et al., 1998), until now only
few applications of Genetic Algorithms have been reported for varieties of the PDPTW.
A Genetic Algorithm for the Dial-a-ride Problem (DARP) is presented by Potter and Bos-
somaier (1995). Since the DARP does not comprise time windows, it is a simplified special
case of the PDPTW. In this approach, the grouping and the routing part of the problem are
solved separately: On the upper level, a Genetic Algorithm is responsible for assigning trans-
portation requests to vehicles, whereas on the lower level a second Genetic Algorithm for each
vehicle determines a route covering the requests assigned to this vehicle.
Jih and Hsu (1999) have developed a hybrid Genetic Algorithm for solving the single-
vehicle PDPTW. The initial population contains a set of routes generated by a dynamic pro-
gramming algorithm which is aborted prior to completion.
Schönberger et al. (2003) propose a Genetic Algorithm for solving the so-called Pickup
and Delivery Selection Problem which extends the PDPTW by the decision of acceptance or
rejection of transportation requests. The goal is to maximize overall profit. To this end, a reve-
 5

nue value is assigned to each request, so this problem significantly differs from the PDPTW
considered in this contribution.
To our knowledge, up to now only Jung and Haghani (2000) have reported a Genetic Al-
gorithm for solving the multi-vehicle PDPTW. In contrast to the PDPTW treated by this paper,
the authors consider so-called soft time windows, i.e. the violation of time window constraints
leads to penalty cost but does not compromise feasibility of a solution.

3 A Grouping Genetic Algorithm for solving the PDPTW

As shown by the literature review, only a few approaches exist which apply Genetic Algo-
rithms to variants of the PDPTW, most of them treating simplified special cases of the
PDPTW. Probably the major reason why Genetic Algorithms for solving the PDPTW are rare
is the fact that it is very difficult to find an appropriate genetic representation for this complex
problem.
Thus, it is an interesting task to verify the appropriateness of Genetic Algorithms for this
problem class. In particular, Genetic Algorithms appear attractive because of their population-
based approach, which allows the parallel exploration of different paths through the solution
space of a problem. In addition, Genetic Algorithms are considered to be robust and easy to
implement (Rayward-Smith, 1994).
In this section, a Genetic Algorithm for solving the PDPTW is proposed. Because it em-
ploys an adaptation of the group-oriented genetic encoding introduced by Falkenauer (1998), it
is denoted as a Grouping Genetic Algorithm (GGA). This section gives a top-down description
of the proposed GGA: First, the basic search scheme of the GGA is outlined. Second, the
group-oriented genetic encoding for PDPTW solutions is explained. Third, the genetic opera-
tors are described in detail. Finally, a rough characterization of the embedded insertion algo-
rithm that is used at several stages of the search is given.

3.1 Genetic search scheme

The concept of Genetic Algorithms was first introduced by Holland (1975). Systematic
treatments of Genetic Algorithms as a class of optimization algorithms can be found in Gold-
berg (1989) and Davis (1991). Figure 1 describes the general search scheme of the proposed
GGA.
 6

initialize population P and evaluate all individuals in P ;

while not (termination criterion is met)
select a pair of individuals x , y from P as parents;
generate two children x′ , y′ applying the crossover operator to x and y with
probability p cross ;
generate two modified children x′′ , y′′ applying the mutation operator to x′ and y′ ,
respectively, with probability p mut ;
evaluate x′′ and y′′ , insert them into P and in turn remove the two worst individuals
from P ;
return best individual from P as solution.

Figure 1: General search scheme of the proposed GGA.

An initial population is generated using the embedded insertion heuristic described below. The
population size, n pop , is a parameter of the GGA. Unlike the classical GA, which employs
generational replacement, the population management is done following the steady-state ap-
proach without duplicates (Syswerda, 1989; Whitley, 1989). According to this approach, each
newly generated pair of offspring is inserted immediately into the current population where it
replaces the two worst individuals. This incremental approach ensures the survival of the best
solution over the whole search and prevents the occurrence of duplicate individuals. Thus, it
often reveals higher performance than generational replacement (Davis, 1991). In order to
detect duplicates, a simple comparison of objective values proved satisfactory for the proposed
GGA, so it was preferred over any other time consuming procedure seeking genotypical or
phenotypical differences between individuals. Note that crossover and mutation are employed
in a sequential fashion. The crossover operator is applied to each selected pair of parent chro-
mosomes with probability pcross , whereas the mutation operator is applied to each offspring
with probability p mut . Both pcross and p mut are parameters of the GGA. If the crossover
operator is not applied according to its execution probability pcross , the children are simply
clones of their parents. Similarly, if the mutation operator is not applied, the offspring leave the
mutation operator unchanged. The search terminates after a given total number, nmax , of
individuals has been generated without improvement but no later than after a given maximum
number, n max , of generated individuals has been reached.
The adaptation of the described genetic search scheme to the problem at hand involves
the following components of the algorithm:
• The genetic encoding, i.e. the way solutions to the PDPTW are represented by chromo-
somes (strings);
• the genetic operators, i.e. selection, crossover and mutation;
 7

• the embedded heuristic, i.e. the subordinate heuristic procedure that is employed by the
GGA in order to generate an initial population and to produce feasible offspring.

The design decisions concerning these components are dealt with in the following sections.

3.2 Group-oriented genetic encoding for the PDPTW

The way in which solutions to the problem at hand are encoded is crucial for the success of any
Genetic Algorithm. One difficulty of finding a good genetic encoding for solutions to the
PDPTW is the fact that the PDPTW – like the VRPTW – is a combination of two interdepen-
dent subproblems: On the one hand, the subproblem of clustering requests and assigning them
to a vehicle has to be solved (clustering or grouping problem). On the other hand, for each
cluster of requests, a feasible route has to be specified in which the corresponding pickup and
delivery nodes are visited (ordering or routing problem). It is not obvious how these two
aspects of a solution can be represented simultaneously by a homogeneous encoding such as
the standard binary representation of the classic Genetic Algorithm.
For pure routing problems like the TSP sometimes a representation is chosen which en-
codes the sequence in which the nodes are visited as a permutation of all locations. A number
of specialized crossover operators that preserve the permutation property of the offspring have
been developed. However, such an encoding brings about some troubles when being applied to
highly constrained multi-vehicle routing problems like the PDPTW:
• Because only the routing part of the PDPTW can be represented by such an encoding, the
grouping part must be covered otherwise, e.g. by external heuristics.
• Due to capacity, precedence and time window constraints a large portion of all possible
permutations will turn out to be infeasible solutions. Thus, special mechanisms have to be
implemented in order to maintain feasibility of the generated solutions.

Among the Genetic Algorithms mentioned in the literature review, Potter and Bossomaier
(1995), Jih and Hsu (1999) as well as Schönberger et al. (2003) use such a permuation-based
encoding.
Jung and Haghani (2000) used a so-called random key representation in order to simulta-
neously encode both the routing and the grouping part of the PDPTW. Under this encoding, a
random four-digit key number is assigned to each location. The first digit denotes the index of
the vehicle visiting that location whereas the last three digits represent a sort key to determine
the position of that location within the route. A solution is decoded by sorting all locations in
ascending order of their key numbers. When generating the chromosome, it is assured that the
value of the pickup location is always less than the value of the corresponding delivery loca-
tion. In addition, the pickup and delivery nodes for the same request always have the same
 8

number in the first digit. Though, many of the individuals generated are infeasible with regard
to capacity and time restrictions and therefore have to be discarded.
Alternatively, a representation could be chosen which encodes a solution as a permutation
of all requests. At the phenotype level, a chromosome of this kind can be decoded by a heuris-
tic which constructs a solution considering the requests in the order given by the permutation.
In the GA literature this rather indirect encoding is quite popular. Applications to the VRP and
the VRPTW are reported for example by Kopfer et al. (1994) and Blanton and Wainwright
(1993), respectively. To our knowledge, this order-based representation has not yet been ap-
plied to the PDPTW.
Nevertheless, the order-based representation with a heuristic decoder has some important
drawbacks, especially for problems with a strong grouping component (see Falkenauer, 1998):
• Identical solutions are often represented by a large number of different permutations. This
redundancy obviously weakens the efficiency of the genetic search.
• The phenotypical meaning of a gene strongly depends on its genotypical context, e.g. the
genes that precede it in the permutation. Most order-based crossover operators are context-
insensitive and do not take into account such contextual information during recombination.
Thus, in many cases, even small modifications in a chromosome, e.g. by a crossover
operator, will cause the offspring to have almost no phenotypical similarities to its parents.
As a consequence, it is hard for the GA to sample meaningful building blocks.
• Due to the strong context-dependency of the genes, good schemata are long, which in turn
increases the probability of them being destroyed by a syntactic crossover operator. This
severely obstructs the search progress of the GA.

There is evidence that the disadvantages of an order-based encoding mentioned above will also
apply to the PDPTW. Furthermore, the grouping part of the PDPTW obviously is of great
importance for the overall quality of a solution to the PDPTW (Savelsbergh and Sol, 1998). In
particular, because both the time windows and the precedence constraints considerably restrict
the number of routing alternatives for a given allocation of requests, the grouping aspect turns
out to be dominant over the routing aspect in many cases.
Because of this dominance of the grouping part of the PDPTW, and in order to avoid the
drawbacks of an order-based encoding, a representation of solutions to the PDPTW is proposed
which refers to the encoding principles of the so-called Grouping Genetic Algorithm (GGA)
introduced by Falkenauer (1998). In this representation, each gene represents a group of ob-
jects instead of a single object. Thus, the groups are the building blocks that are sampled and
recombined by the genetic operators to produce solutions of potentially higher quality. Falke-
nauer originally developed his GGA for pure grouping problems such as the Bin Packing Prob-
lem (BPP). The following describes how the original encoding was adapted to the PDPTW.
 9

Each gene in a chromosome represents a cluster of all requests that are assigned to a sin-
gle vehicle. The length of a chromosome, i.e. the number of genes is variable and depends on
the number of vehicles required by a given solution. Figure 2 illustrates the encoding of a solu-
tion by means of a simple example with seven requests that are carried out by three vehicles.

2- 1- 3-

1
7- 5-
4+ -
4
-
6 6+ 7+ solution
2+ (phenotype)
8
3 0
3+
1+ 5+
0 0

4 1 3 cluster
1 3 8 5
6 (genotype)
2 7

0 - home location
i + - pickup location
i - - delivery location
i - request

Figure 2: Group-oriented encoding (example).

By representing a partitioning of all transportation requests, a chromosome under this encoding

only covers the grouping aspect of a PDPTW solution. As a consequence, the remaining rou-
ting aspect of a solution has to be added while decoding the chromosome. This is done by asso-
ciating each chromosome with separate data structures containing the routing information for
each gene. Since this information is hidden from the GGA, it cannot be manipulated directly by
the genetic operators. Instead, the routes are constructed and maintained using a separate heu-
ristic according to the grouping information encoded in the chromosome.

3.3 Genetic operators

In the proposed GGA, the binary tournament selection mechanism (see e.g. Goldberg et al.,
1990) is applied as the selection operator. In this scheme, two individuals are chosen randomly
and the one which represents the better solution is selected as the first parent. To obtain the
second parent, the procedure is repeated. An important advantage of tournament selection is its
low time complexity compared to the classical “roulette wheel” selection scheme. Furthermore,
because a simple comparison of the objective values is sufficient, tournament selection does
not require the calculation of special fitness values.
 10

The crossover operator for the proposed GGA is an adaptation of the general group-
oriented crossover scheme presented in Falkenauer (1998) and proceeds in five steps, as illus-
trated in Figure 3.

1. Specify crossing sections

4 3 1
parent 1 3 2 8 1 8
7 6 5

4 1 2 3
parent 2 2 1 5 8
8 5
6 7

2. Insert groups

4 1 2 3 1
child 1 3 2 1 5 8 1
8
7 6 7 6 5

3. Clean up chromosome

1 2 3
3 4 1 5 8 8 5
6 7

4. Re-insert unassigned requests

4 1 2 3
3 1 5 5 8
8 6 7 5

Figure 3: The group-oriented crossover operator.

1. Specify a crossing section, i.e. a coherent segment of clusters, by randomly selecting two
crossing points in each of the two parent chromosomes.
2. Insert the clusters in the crossing section of the second parent at the first crossing point in
the chromosome of the first parent. The respective routes are directly adopted from the sec-
ond parent without reconstruction. As a result of this operation a number of requests may
occur twice (requests 1, 2, 6 and 7 in Figure 3). Moreover, two clusters may be assigned to
the same vehicle (vehicle 1 in Figure 3).
 11

3. To resolve the conflicts mentioned above, remove all clusters originally belonging to the
first parent if they refer to a vehicle that is already allocated by any of the newly inserted
clusters. If afterwards some requests still occur twice, eliminate them from those clusters
that originally belong to the first parent. At the phenotype level, all eliminations are repro-
duced immediately in the routes appendant to the changed clusters. Note that the clusters
and the respective routes imported from the second parent are left unchanged. After this
step, some requests may remain unassigned (requests 5 and 8 in Figure 3).
4. Reinsert the unassigned requests in random order into the individual applying the insertion
heuristic. This may require the allocation of an additional vehicle in order to assure feasibil-
ity. As a result of this step, a complete offspring is obtained.
5. Generate the second offspring by repeating steps 2 through 4 with reversed roles of the par-
ents.

The group-oriented mutation operator for the PDPTW works as follows:

1. Select a cluster at random in the individual.
2. Eliminate this cluster from the chromosome and remove the associated route from the phe-
notype.
3. Re-insert all removed requests into the individual by means of the insertion heuristic, allo-
cating a new vehicle if necessary to maintain feasibility.

3.4 Embedded insertion heuristic

At several stages during search, the GGA employs a subordinate insertion heuristic while
constructing and modifying solutions. The general procedure of this heuristic can be described
as follows: For each request to be inserted, all feasible insertions in all existing routes of the
current (partial) solution are examined. To this end, all possible insertion positions for the
pickup and the delivery node in a route are tested taking into consideration precedence, capa-
city and time constraints. Additionally, a new vehicle is allocated and a new route for this
vehicle is tentatively initialized with the current request. Among all feasible insertions, the one
that causes minimal additional cost is selected and implemented.
In order to generate an initial solution by means of this procedure, all transportation re-
quests are considered in random order, starting with a single, empty vehicle. After all transpor-
tation requests have been inserted, a feasible initial solution is available. The procedure is re-
peated until an initial population of the desired size is obtained.
In order to re-adjust the routing part of a solution due to grouping modifications caused
by the genetic operators, the insertion procedure is also used. Instead of constructing a new
solution from scratch every time the grouping is modified by crossover or mutation, each re-
quest that is subject to regrouping is deleted from its original route and reinserted without dis-
 12

turbing the relative order in which the remaining nodes are visited. Thus, the original routing
decisions in the unchanged parts of a solution are preserved.

4 Computational Study
This section reports the results of the proposed GGA for a number of benchmark problems
when compared to previous metaheuristics for the PDPTW. First, the benchmark data sets are
described, then the numerical results are discussed.

4.1 Benchmark Data Sets

It seems that up to now only two benchmark data sets for the PDPTW for which results have
been published in the literature are publicly available:
1. The nine 100-node-instances provided by Nanry and Barnes (2000), in the remainder re-
ferred to as problem set 1. These instances are available via ftp from www.me.utexas.edu/
pub/barnes. Results for these instances are also reported by Li and Lim (2001) and Lau and
Liang (2001).
2. The 56 100-node-instances provided by Li and Lim (2001), in the remainder referred to as
problem set 2. Problem set 2 can be downloaded from www.comp.nus.edu.sg/~lihb/
Pdptw.html.

In these problem sets, each PDPTW instance was derived from a certain instance out of
Solomon’s 100-customer VRPTW problems by pairing up customers within the routes of a
given reference solution. Supplementary dummy nodes were introduced for coupling purposes
if an odd number of customers is present in a route of the original solution. So each instance
has at least 100 nodes, which is equivalent to at least 50 requests. As Li and Lim (2001), as
well as Lau and Liang (2001) point out, it seems that Nanry and Barnes paired customers
directly adjacent to each other in the reference solution while generating most of their problem
instances. In contrast to this, Li and Lim used a random approach.
Nanry and Barnes restricted themselves to Solomon instances for which optimal solutions
are known. They only generated instances from Solomon’s C1 class. In all problems of this
class, the customers are spatially clustered and the problems have a short scheduling horizon.
For example, the instance denoted as nc101 is derived from Solomon’s VRPTW instance c101.
For consistency reasons, minor adjustments were necessary to some of the instances of Nanry
and Barnes (2001) because, for some requests, the pickup and delivery quantities were not
equal.
In order to generate PDPTW test problems from all Solomon instances, Li and Lim refer
to near-optimal solutions in cases where the optimal solution is unknown. Following the classi-
fication of Solomon, they organized their instances into six classes, denoted as LC1, LC2, LR1,
 13

LR2 LRC1 and LRC2, indicating the respective spatial distribution of the nodes and the length
of the scheduling horizon. In LC1 and LC2 problems, nodes are clustered, whereas in LR1 and
LR2 problems, they are randomly distributed. The nodes in LRC1 and LRC2 instances are par-
tially clustered and partially randomly distributed. While the instances in LC1, LR1 and LRC1
have a short scheduling horizon, the horizon is longer in instances of classes LC2, LR2 and
LRC2.

4.2 Numerical Results

The GGA was implemented in JAVA and evaluated using problem set 1 and problem set 2. All
tests were carried out on a PC (Pentium 4 processor, 2 GHz). The GGA was run 30 times on
each instance under the objective of minimizing total travel distance. Euclidean distances and
travel times were calculated to two decimal places as is standard practice in the recent VRPTW
literature. The parameter settings for the GGA, which were determined in a small number of
preliminary experiments, are shown in table 1. The values were chosen mainly with the idea of
performing a thorough search but also to keep computation times reasonable.

Table 1: GGA parameter values

Parameter value description
n pop 200 population size

pcross 1.0 crossover probability

p mut 0.5 mutation probability

n max 15,000 max. number of individuals

nmax 3,000 max. number of individuals

without improvement

In the remainder, the results obtained by the GGA on problem set 1 will be compared to the
results reported by Nanry and Barnes (2000), Li and Lim (2001) and Lau and Liang (2001).
Since no other results are published for problem set 2, the GGA results for these instances are
compared to the results of Li and Lim (2001) only. However, it is difficult to directly compare
the results of the different approaches.
One important problem is that the approaches apply slightly different objective criteria.
Nanry and Barnes (2000) conducted their experiments under the objective of minimizing the
sum of total travel time and total service time. Since vehicle speed is set to 1 and total service
time is a constant if all nodes are visited, this objective is basically equivalent to minimizing
total travel distance, which was chosen as the objective for the GGA. In contrast, Li and Lim
(2001) used a weighted sum of the number of vehicles, total travel distance, schedule duration
and total waiting time as the objective function. Similarly, the objective function considered by
 14

Lau and Liang (2001) is a weighted sum of the number of vehicles and total travel distance.
Furthermore, the approaches differ in the way euclidean distances are calculated. Even though
it is not explicitly stated in the paper, the objective values of the optimal VRPTW solutions
used as a benchmark reveal that in Nanry and Barnes (2000), distance is rounded or truncated
to one decimal position. In contrast, it seems that Lim and Li (2001) as well as Lau and Liang
(2001) calculate distance to at least two decimal points. Therefore, even identical solutions
may lead to significantly different objective values. Moreover, only best results are reported in
the relevant literature, even if the proposed methods have stochastic components, as in the ap-
proaches of Li and Lim (2001) and Lau and Liang (2001). The overall number of independent
runs per instance and average solution quality, which are important pieces of information when
judging a stochastic method, are not disclosed. Finally, due to different computer platforms, a
comparison of computing times is difficult.
The results obtained by the GGA for problem set 1 and problem set 2 are listed in table 2
and table 3, respectively. In these tables, td denotes total travel distance, nv the number of
vehicles, and ct the computing time in seconds. Since Lau and Liang (2000) did not report a
solution to nc102 nor any computing times, the respective cells in table 2 are empty. For the
GGA, aside from the values of the respective best solution ( td best and nv best ), average solution
quality and average computing time is reported for each instance ( td avg , nv avg , and ct avg ). Un-
der the assumption that the optimal VRPTW solutions remain very good, if not optimal solu-
tions to the derived PDPTW instances, each of them provides a high quality upper bound for
the total travel distance of the corresponding PDPTW instance. For this reason, table 2 also
shows the optimal values for the corresponding VRPTW instances of problem set 1 (columns
two and five). In order to allow a careful comparison of the approaches despite different dis-
tance calculations, the results of Nanry and Barnes are compared to the optimal VRPTW solu-
tions of Desrochers et al. (1992) and Kohl and Madsen (1997), which were calculated using
distance truncation after one decimal place (Opt A). When calculated to two decimal places,
the objective values of the optimal VRPTW solutions are slightly different (Opt B, taken from
Homberger, 2000). These values should be taken into account when comparing the results of
the rest of the approaches, including the GGA.
The results presented in table 2 can be summarized as follows:
• For problem nc103, the GGA, like the competing methods, fails to find a solution of the
same quality as the corresponding optimal VRPTW solution (Opt B). The relative error is
0.5 percent. Only the solution reported by Nanry and Barnes for this instance has a lower
relative error (0.4 percent).
• For all other problems, the GGA finds solutions of identical or better quality in terms of
total travel distance when compared to the corresponding optimal VRPTW solutions. Since
the average computing times are less than 2.5 minutes even in the worst case, this solution
quality is achieved at reasonable computational cost.
 15

Table 2: Results for problem set 1

Opt. Nanry/Barnes Opt. Li/Lim Lau/Liang
Inst. GGA
A (2000) B (2001) (2001)
td td ct td td ct td ct tdbest tdavg ctavg
nv nv nv nv nv nvbest nvavg

nc101 827.3 827.3 0.25 828.94 828.9 35 828.9 – 828.94 828.94 64.93
10 10 10 10 10 10 10
nc102 827.3 827.3 1.91 828.94 828.9 130 – – 828.94 828.94 83.1
10 10 10 10 10 10
nc103 826.3 829.9 16.50 828.06 831.9 156 837.7 – 831.87 831.87 117.13
10 10 10 10 10 10 10
nc104 822.9 834.7 328.87 824.78 869.5 1539 989.7 – 816.74 816.74 143.67
10 10 10 9 9 10 10
nc105 827.3 827.3 0.51 828.94 828.9 25 829.8 – 828.94 828.94 69.83
10 10 10 10 10 10 10
nc106 827.3 827.3 0.54 828.94 828.9 31 828.9 – 828.94 828.94 74.73
10 10 10 10 10 10 10
nc107 827.3 826.1 9.36 828.94 827.8 22 828.9 – 827.82 827.82 74.10
10 10 10 10 10 10 10
nc108 827.3 826.1 31.49 828.94 827.8 39 826.1 – 827.82 829.11 84.9
10 10 10 10 10 10 10
nc109 827.3 827.3 184.76 828.94 827.8 85 828.9 – 827.82 827.82 116.87
10 10 10 10 10 10 10

• For problem nc104, the GGA has found a new best solution with regard to total travel
distance. Interestingly, total travel distance in this case is significantly less compared to the
corresponding optimal VRPTW solution. This also holds for three further instances (nc107,
nc108 and nc109). Due to load modifications at the customer nodes during transformation
into PDPTW instances, the capacity constraint obviously has lost its original tightness.
Thus, the optimal solutions for the corresponding VRPTW are no longer optimal for the
generated PDPTW instances. Note that for problem nc104, Li and Lim as well as Lau and
Liang report solutions with greater travel distance but one vehicle saved.
• An interesting observation is the robustness of the proposed GGA with regard to problem
set 1, since except for instance nc108, the GGA finds its best solution to each problem
consistently in all 30 runs.

In order to facilitate comparison, in table 3, the respective best values with regard to travel
distance and number of vehicles for each instance are highlighted as bold figures. From table 3
the following conclusions can be drawn:
• In comparison to the method of Li and Lim (2001), the best results of the GGA are better
with respect to total travel distance in 15 cases. On the other hand, in 9 cases the method of
Li and Lim yields solutions with less travel distance than the GGA.
 16

Table 3: Results for problem set 2

Inst. Li and Lim (2001) GGA
best best
td nv ct td nv tdavg nvavg ctavg
lc101 828.94 10 33 828.94 10 828.94 10.00 55.03
lc102 828.94 10 71 828.94 10 828.94 10.00 68.67
lc103 827.86 10 191 827.86 10 827.86 10.00 85.33
lc104 861.95 9 1254 818.60 10 818.67 10.00 139.90
lc105 828.94 10 47 828.94 10 828.94 10.00 58.33
lc106 828.94 10 43 828.94 10 828.94 10.00 63.60
lc107 828.94 10 54 828.94 10 828.94 10.00 65.97
lc108 826.44 10 82 826.44 10 826.44 10.00 81.00
lc109 827.82 10 255 827.82 10 827.82 10.00 118.10
lc201 591.56 3 27 591.56 3 591.56 3.00 59.27
lc202 591.56 3 94 591.56 3 591.56 3.00 115.50
lc203 585.56 3 145 591.17 3 591.17 3.00 191.67
lc204 591.17 3 746 590.60 3 620.38 3.00 346.13
lc205 588.88 3 190 588.88 3 588.88 3.00 94.07
lc206 588.49 3 88 588.49 3 588.49 3.00 124.20
lc207 588.29 3 102 588.29 3 588.35 3.00 137.03
lc208 588.32 3 178 588.32 3 588.75 3.00 141.93
lr101 1650.78 19 87 1650.80 19 1650.80 19.00 67.00
lr102 1487.57 17 1168 1487.57 17 1488.74 17.00 89.53
lr103 1292.68 13 169 1292.68 13 1292.67 13.00 82.87
lr104 1013.39 9 459 1013.99 9 1037.16 9.43 136.83
lr105 1377.11 14 69 1377.11 14 1377.11 14.00 70.97
lr106 1252.62 12 87 1252.62 12 1252.63 12.00 81.33
lr107 1111.31 10 287 1111.31 10 1112.26 10.03 95.67
lr108 968.97 9 415 968.97 9 969.02 9.00 102.50
lr109 1239.96 11 348 1208.96 11 1233.99 12.00 103.80
lr110 1159.35 10 547 1165.83 11 1176.20 11.13 142.13
lr111 1108.90 10 179 1108.90 10 1113.60 10.20 105.53
lr112 1003.77 9 638 1003.77 9 1040.34 10.27 168.70
lr201 1263.84 4 193 1253.23 4 1260.41 4.20 103.43
lr202 1197.67 3 885 1197.67 3 1255.78 4.03 234.50
lr203 949.40 3 1950 952.29 3 962.82 3.00 381.50
lr204 849.05 2 2655 849.05 2 879.01 2.27 486.50
lr205 1054.02 3 585 1054.02 3 1078.84 3.03 166.37
lr206 931.63 3 747 931.63 3 941.67 3.00 259.03
lr207 903.056 2 1594 903.60 2 927.58 2.43 418.77
lr208 734.85 2 3572 736.00 2 760.95 2.03 531.07
lr209 937.05 3 2773 932.43 3 955.96 3.07 236.90
lr210 964.22 3 1482 964.22 3 972.34 3.00 286.13
lr211 927.80 2 4204 888.15 3 897.03 3.07 478.53
lrc101 1708.80 14 119 1703.21 15 1704.05 15.00 76.07
lrc102 1563.55 13 152 1558.07 12 1559.65 12.17 95.53
lrc103 1258.74 11 175 1258.74 11 1260.45 11.00 92.80
lrc104 1128.40 10 202 1128.40 10 1129.32 10.00 108.23
lrc105 1637.62 13 179 1637.62 13 1639.87 13.10 88.17
lrc106 1425.53 11 459 1424.73 11 1441.82 11.77 95.63
lrc107 1230.15 11 154 1230.14 11 1237.91 11.07 97.73
lrc108 1147.97 10 650 1147.43 10 1159.82 10.40 111.67
lrc201 1468.96 4 266 1407.21 4 1438.12 4.77 101.47
lrc202 1374.27 3 987 1385.25 4 1400.11 3.97 161.70
lrc203 1089.07 3 1605 1093.89 4 1114.40 4.00 268.17
lrc204 827.78 3 3634 818.66 3 838.55 3.00 457.30
lrc205 1302.20 4 639 1302.20 4 1305.37 4.00 147.20
lrc206 1162.91 3 445 1159.03 3 1176.90 3.63 139.57
lrc207 1424.60 3 607 1062.05 3 1070.56 3.03 218.97
lrc208 852.76 3 4106 852.76 3 882.78 3.20 322.13
 17

• In 11 cases the GGA finds solutions with less travel distance without an increase in the
number of vehicles reported by Li and Lim. These new best results are printed in bold
italics in table 3. For problem lrc102, the GGA finds solutions that have less total travel
distance and a smaller number of vehicles than the best solutions of Li and Lim. In 6 cases,
Li and Lim report solutions with a smaller number of vehicles than the GGA.
• Even though the GGA fails to find its best solution for each instance of problem set 2
throughout all runs, the GGA turns out to be reasonably robust: for almost 90 percent of all
instances, a standard deviation of total travel distance of less than 2 percent was observed.
• In order to solve the instances of problem set 2, the GGA requires moderate average
computational times that range from less than one minute to less than nine minutes.

5 Conclusions
In this paper, a Grouping Genetic Algorithm (GGA) for the Pickup and Delivery Problem with
Time Windows (PDPTW) has been presented. To our knowledge this is the first time a Genetic
Algorithm has been applied to the multi-vehicle PDPTW with hard time window constraints.
Furthermore, it appears to be the first published application of the group-oriented genetic
encoding of Falkenauer (1998) to a generalization of the Vehicle Routing Problem. The GGA
was tested using two publicly available sets of benchmark problems for the PDPTW. The expe-
rimental results have demonstrated that the GGA is able to find high quality solutions when
compared to previous metaheuristic methods for solving the PDPTW. The overall findings
seem to justify the employment of Genetic Algorithms in general, as well as the group-oriented
encoding in particular, as suitable techniques for solving the PDPTW.
 18

References
BERGER, J.; SALOIS, M.; BEGIN, R. (1998): A Hybrid Genetic Algorithm for the Vehicle
Routing Problem with Time Windows. In: Mercer, R.E.; Neufeld, E. (eds.), Advances in
Artificial Intelligence – Proceedings of the 12th Biennial Conference of the Canadian
Society for Computational Studies of Intelligence, Springer, Berlin, pp. 114–127.
BLANTON, J.L.; WAINWRIGHT, R.L. (1993): Multiple Vehicle Routing with Time und Capacity
Constraints using Genetic Algorithms. In: Forrest, S. (ed.), Proceedings of the Fifth Inter-
national Conference on Genetic Algorithms, Morgan Kaufmann, San Mateo/California,
452–459.
BODIN, L.; SEXTON, T. (1986): The Multi-Vehicle Subscriber Dial-a-Ride Problem. TIMS
Studies in the Management Sciences 26, pp. 73–86.
DAVIS, L. (ed.) (1991): Handbook of Genetic Algorithms. Van Nostrand Reinhold, New York.
DESROCHERS, M.; DESROSIERS, J.; SOLOMON, M.M. (1992): A New Optimization Algorithm
for the Vehicle Routing Problem with Time Windows. Operations Research 40, pp. 342–
354.
DESROSIERS, J.; DUMAS Y.; SOUMIS, F. (1986): A Dynamic Programming Solution of the
Large-Scale Single-Vehicle Dial-a-Ride Problem with Time Windows. The American
Journal of Mathematical und Management Sciences 6, pp. 301–325.
DESROSIERS, J.; DUMAS, Y.; SOUMIS, F. (1988): The Multiple Vehicle Dial-a-Ride Problem. In:
Daduna, J.R.; Wren, A. (eds.), Computer-Aided Transit Scheduling: Proceedings of the
Fourth International Workshop on Computer-Aided Scheduling of Public Transport, Sprin-
ger, Berlin, pp. 15–27.
DUMAS, Y.; DESROSIERS, J.; SOUMIS F. (1991): The Pickup und Delivery Problem with Time
Windows. European Journal of Operational Research 54, pp. 7–22.
FALKENAUER, E. (1998): Genetic Algorithms und Grouping Problems, Wiley, Chichester.
GENDREAU, M.; GUERTIN, F.; POTVIN, J.-Y.; SÉGUIN, R. (1998): Neighborhood Search Heu-
ristics for a Dynamic Vehicle Dispatching Problem with Pick-ups und Deliveries. Tech-
nical Report CRT-98-10, Centre de recherche sur les transports, Université de Montréal,
Montréal.
GOLDBERG, D.E. (1989): Genetic Algorithms in Search, Optimization, und Machine Learning.
Addison-Wesley, Reading, Massachusetts.
GOLDBERG, D.E.; KORB, B.; DEB, K. (1990): Messy Genetic Algorithms: Motivation, Analysis,
und First Results. Complex Systems 3, pp. 493–530.
HOLLAND, J.H. (1975): Adaptation in Natural und Artificial Systems – An Introductory Analy-
sis with Applications to Biology, Control, und Artificial Intelligence. The University of
Michigan Press, Ann Arbor.
 19

HOMBERGER, J. (2000): Verteilt-parallele Metaheuristiken zur Tourenplanung – Lösungs-

verfahren für das Standardproblem mit Zeitfensterrestriktionen, Deutscher Universitäts-
Verlag, Wiesbaden.
IOACHIM, I.; DESROSIERS, J.; DUMAS, Y.; SOLOMON M.M.; VILLENEUVE D. (1995): A Request
Clustering Algorithm for Door-to-Door Handicapped Transportation. Transportation
Science 29, pp. 63–78.
JAW, J.-J.; ODONI, A.R.; PSARAFTIS, H.N.; WILSON, N.H.M. (1986): A Heuristic Algorithm for
the Multi-Vehicle Advance Request Dial-A-Ride Problem with Time Windows. Trans-
portation Research Part B 20, pp. 243–257.
JIH, W.-R.; HSU, Y.-J. (1999): Dynamic Vehicle Routing Using Hybrid Genetic Algorithms. In:
IEEE Computer Society (ed.), Proceedings of the 1999 IEEE International Conference on
Robotics & Automation, IEEE Computer Society, Los Alamitos, California, pp. 453–458.
JUNG, S.; HAGHANI, A. (2000): A Genetic Algorithm for Pick-up and Delivery Problem with
Time Windows. In: Transportation Research Record 1733, Transportation Research Board,
pp. 1–7.
KOHL, N.; MADSEN, O.B.G (1997): An Optimization Algorithm for the Vehicle Routing Prob-
lem with Time Windows Based on Lagrangian Relaxation. Operations Research 45,
pp. 395–406.
KOPFER, H.; PANKRATZ, G.; ERKENS, E. (1994): Die Entwicklung eines hybriden Genetischen
Algorithmus für das Tourenplanungsproblem. OR Spektrum 16, pp. 21–32.
LAU, H.C.; LIANG, Z. (2001): Pickup und Delivery with Time Windows, Algorithms und Test
Case Generation. In: IEEE Computer Society (ed.), Proceedings of the 13th IEEE Inter-
national Conference on Tools with Artificial Intelligence (ICTAI), IEEE Computer
Society, Los Alamitos, California, pp. 333–340.
LENSTRA, J.K.; RINNOY KAN, A.H.G. (1981): Complexity of Vehicle Routing und Scheduling
Problems. Networks 11, pp. 221–227.
LI, H.; LIM, A. (2001): A Metaheuristic for Solving the Pickup und Delivery Problem with
Time Windows. In: IEEE Computer Society (ed.), Pro-ceedings of the 13th IEEE Inter-
national Conference on Tools with Artificial Intelligence (ICTAI), IEEE Computer
Society, Los Alamitos, California, pp. 160–167.
LIN, S.; KERNIGHAN, B. (1973): An Effective Heuristic Algorithm for the Traveling Salesman
Problem. Operations Research 21, pp. 498–516.
MADSEN, O.B.G.; RAVN, H.F.; RYGAARD, J.M. (1995): A Heuristic Algorithm for a Dial-a-
Ride Problem with Time Windows. Multiple Capacities and Multiple Objectives. Annals
of Operations Research 60, pp. 193–208.
NANRY, W.P.; BARNES, J.W. (2000): Solving the Pickup und Delivery Problem with Time
Windows using Reactive Tabu Search. Transportation Research Part B 34, pp. 107–121.
 20

POTTER, T.; BOSSOMAIER, T. (1995): Solving Vehicle Routing Problems with Genetic Algo-
rithms. In: IEEE Computer Society (ed.), Proceedings of the 1995 IEEE Interational Con-
ference on Evolutionary Computing, IEEE Computer Society, Los Alamitos, California,
pp. 788–793.
POTVIN, J.-Y.; BENGIO, S. (1996): The Vehicle Routing Problem with Time Windows – Part II:
Genetic Search. INFORMS Journal on Computing 8, pp. 165–172.
PSARAFTIS, H.N. (1980): A Dynamic Programming Solution to the Single Vehicle Many-to-
Many Immediate Request Dial-a-Ride Problem. Transportation Science 14, pp. 130–154.
PSARAFTIS, H.N. (1983a): An Exact Algorithm for the Single Vehicle Many-to-Many Dial-A-
Ride Problem with Time Windows. Transportation Science 17, pp. 351–357.
PSARAFTIS, H.N. (1983b): k-Interchange Procedures for Local Search in a Precedence-
Constrained Routing Problem. European Journal of Operational Research 13, pp. 391–402.
RAYWARD-SMITH, V.J. (1994): A Unified Approach to Tabu Search, Simulated Annealing und
Genetic Algorithms. In: Unicom Seminars Ltd. (ed.), Adaptive Computing und Informa-
tion Processing, Unicom Seminars Ltd., London, Vol. I, pp. 55–78.
REGO, C.; ROUCAIROL, C. (1996): A Parallel Tabu Search Algorithm Using Ejection Chains for
the Vehicle Routing Problem. In: Osman, I.H.; Kelly, J.P. (eds.), Meta-Heuristics: Theory
and Applications, Kluwer, Norwell et al., pp. 661–675
SAVELSBERGH, M.W.P.; SOL, M. (1995): The General Pickup und Delivery Problem. Trans-
portation Science 29, pp. 17–29.
SAVELSBERGH, M.W.P.; SOL, M. (1998): DRIVE: Dynamic Routing of Independent Vehicles.
Operations Research 46, pp. 474–490.
SCHÖNBERGER, J.; KOPFER, H.; MATTFELD, D.C. (2003): A Combined Approach to Solve the
Pickup und Delivery Selection Problem. In: Leopold-Wildburger, U.; Rendl, F.; Wäscher,
G. (eds.), Operations Research Proceedings 2002, Springer, Heidelberg et al., pp. 150–155.
SOLOMON, M. (1987): Algorithms for the Vehicle Routing und Scheduling Problem with Time
Window Constraints. Operations Research 35 (2), pp. 254–265.
SYSWERDA, G. (1989): Uniform Crossover in Genetic Algorithms. In: Shaffer, J.D. (ed.), Pro-
ceedings of the Third International Conference on Genetic Algorithms, Morgan Kauf-
mann, San Mateo/California, pp. 2–9.
THANGIAH, S.R. (1995): Vehicle Routing with Time Windows using Genetic Algorithms. In:
Chambers, L. (ed.), Practical Handbook of Genetic Algorithms 2 – New Frontiers. CRC
Press, Boca Raton, pp. 253–277.
TOTH, P.; VIGO, D. (1996): Fast Local Search Algorithms for the Handicapped Persons Trans-
portation Problem. In: Osman, I.H.; Kelly, J.P. (eds.), Meta-Heuristics: Theory und appli-
cations, Kluwer, Norwell, pp. 677–690.
 21

VAN DER BRUGGEN, L.J.J.; LENSTRA, J.K.; SCHUUR, P.C. (1993): Variable Depth Search for
the Single-Vehicle Pickup und Delivery Problem with Time Windows. Transportation
Science 27, pp. 298–311.
WHITLEY, L.D. (1989): The GENITOR Algorithm und Selection Pressure: Why Rank-Based
Allocation of Reproductive Trials is Best. In: Shaffer, J.D. (ed.), Proceedings of the Third
International Conference on Genetic Algorithms, Morgan Kaufmann, San Mateo/Cali-
fornia, pp. 116–121.
 22

Die Diskussionspapiere ab Nr. 183 (1992) bis heute, können Sie im Internet unter
http://www.fernuni-hagen.de/FBWIWI/ einsehen und zum Teil downloaden.
Die Titel der Diskussionspapiere von Nr. 1 (1975) bis 182 (1991) können bei Bedarf im Fachbereich Wirt-
schaftswissenschaft angefordert werden:
FernUniversität
z.Hd. Frau Huber oder Frau Mette,
Postfach 940
58084 Hagen
Die Diskussionspapiere selber erhalten Sie nur in den Bibliotheken.

Nr. Jahr Titel Autor/en

322 2001 Spreading Currency Crises: The Role of Economic Interdependence Berger, Wolfram
Wagner, Helmut
323 2002 Planung des Fahrzeugumschlags in einem Seehafen- Fischer, Torsten
Automobilterminal mittels eines Multi-Agenten-Systems Gehring, Hermann
324 2002 A parallel tabu search algorithm for solving the container loading Bortfeldt, Andreas
problem Gehring, Hermann
Mack, Daniel
325 2002 Die Wahrheit entscheidungstheoretischer Maximen zur Lösung von Mus, Gerold
Individualkonflikten – Unsicherheitssituationen
326 2002 Zur Abbildungsgenauigkeit des Gini-Koeffizienten bei relativer Steinrücke, Martin
wirtschaftlicher Konzentration
327 2002 Entscheidungsunterstützung bilateraler Verhandlungen über Steinrücke, Martin
Auftragsproduktionen – eine Analyse aus Anbietersicht
328 2002 Die Relevanz von Marktzinssätzen für die Investitionsbeurteilung – Terstege, Udo
zugleich eine Einordnung der Diskussion um die Marktzinsmethode
329 2002 Evaluating representatives, parliament-like, and cabinet-like represen- Tangian, Andranik S.
tative bodies with application to German parliament elections 2002
330 2002 Konzernabschluss und Ausschüttungsregelung im Konzern. Ein Hinz, Michael
Beitrag zur Frage der Eignung des Konzernabschlusses als
Ausschüttungsbemessungsinstrument
331 2002 Theoretische Grundlagen der Gründungsfinanzierung Bitz, Michael
332 2003 Historical background of the mathematical theory of democracy Tangian, Andranik S.
333 2003 MCDM-applications of the mathematical theory of democracy: Tangian, Andranik S.
choosing travel destinations, preventing traffic jams, and predicting
stock exchange trends
334 2003 Sprachregelungen für Kundenkontaktmitarbeiter – Möglichkeiten und Fließ, Sabine
Grenzen Möller, Sabine
Momma, Sabine B.
335 2003 A Non-cooperative Foundation of Core-Stability in Positive Finus, Michael
Externality NTU-Coalition Games Rundshagen, Bianca
336 2003 Combinatorial and Probabilistic Investigation of Arrow’s paradox Tangian, Andranik
337 2003 A Grouping Genetic Algorithm for the Pickup and Delivery Problem Pankratz, Giselher
with Time Windows
338 2003 Planen, Lernen, Optimieren: Beiträge zu Logistik und E-Learning. Bortfeldt, Andreas
Festschrift zum 60. Geburtstag von Hermann Gehring Fischer, Torsten
Homberger, Jörg
Pankratz, Giselher
Strangmeier, Reinhard

View publication stats