Analyzing A Unified Ant System For The VRP and Some of Its Variants
Analyzing A Unified Ant System For The VRP and Some of Its Variants
Analyzing A Unified Ant System For The VRP and Some of Its Variants
1 Introduction
Vehicle routing and Scheduling has been of major interest to the scientific
community for the last 50 years due to the inherent complexity of the asso-
ciated problems. A large number of different approaches, in recent years mostly
meta-heuristics have been proposed for different variants of vehicle routing and
scheduling problems, see further Toth and Vigo [1].
On the other hand, goods distribution and more specifically vehicle rout-
ing represents an important area of an economy. Basically, in every supply chain
transportation occurs between member companies of the chain or from the chain
to final customers. Thus, firms have recognized the need to automatize this pro-
cess and use software to support their distribution process. Part of this software
is an optimization tool for routing and scheduling.
However, while it seems that scientific interest and company interest overlap
in this regard, this first impression is not completely true. Rather, the level of
development of tools in the scientific world is not matched in industry. Most
of the tools used in industry are based on very simple optimization techniques
developed many years ago.
The reasons for this are manyfold. First, it can not be expected that cutting
edge research will instantaneously be transferred to industry. Rather a certain lag
has to be accepted. Second, in the quest for ever more sophisticated optimization
techniques the scientific community has reached a level of specialization that
prevents the tools developed from being applicable to a wide range of problems.
Thus, while these tools solve particular problem scenarios much better than
the ’old’ simple techniques, they lack the flexibility of the latter approaches.
However, for practical purposes solution quality is not the only criterion for the
choice of an optimization tool.
This gap has been recognized by some researchers. For example Cordeau et
al.[2] present four measures of algorithm performance, namely solution quality,
S. Cagnoni et al. (Eds.): EvoWorkshops 2003, LNCS 2611, pp. 300–310, 2003.
c Springer-Verlag Berlin Heidelberg 2003
Analyzing a Unified Ant System for the VRP and Some of Its Variants 301
speed, flexibility and simplicity. According to their analysis most of the meta-
heuristics, while outperforming simple heuristics with respect to solution quality
can not compete with simple heuristics when it comes to speed or simplicity.
Simplicity is customarily measured in terms of the number of parameters of an
algorithm, while flexibility is concerned with the applicability of an algorithm
to different variants of a basic problem. This applicabilty covers also issues of
robustness, i.e. whether parameter changes are necessary if the problem instance
to be solved changes. On the other hand, a research group in Oslo works on a
formal representation and algorithms for rich vehicle routing problems, which
cover a wide range of problems with different constraints and objectives (c.f. [3]).
In light of these observations, the aim of our paper is to present an Ant
System algorithm that is capable of solving different variants of vehicle routing
problems reasonably well, with little or no adjustments to the implementation.
More specifically, we aim to solve the vehicle routing problem (VRP), the vehicle
routing problem with time windows (VRPTW), the vehicle routing problem with
backhauls (VRPB) and the vehicle routing problem with backhauls and time
windows (VRPBTW). The algorithm is based on the Insertion based AS we
presented in Reimann et al. [4].
The remainder of this paper is organized as follows. In the next section we
describe the characteristics of the different problems we aim to solve. After that
we briefly describe our Unified AS approach. In section 4 we present our rich
computational study, before we conclude in Section 5.
The Ant System approach, originally proposed by Colorni et al. (see e.g. [11])
is based on the behavior of real ants searching for food. Real ants communicate
with each other using an aromatic essence called pheromone, which they leave
on the paths they traverse. In the absence of pheromone trails ants more or less
perform a random walk. However, as soon as they sense a pheromone trail on
a path in their vicinity, they are likely to follow that path, thus reinforcing this
trail. More specifically, if ants at some point sense more than one pheromone trail,
Analyzing a Unified Ant System for the VRP and Some of Its Variants 303
they will choose one of these trails with a probability related to the strenghts
of the existing trails. This idea has first been applied to the TSP, where an ant
located in a city chooses the next city according to the strength of the artificial
trails.
Improved versions of the basic algorithm have been applied to a large num-
ber of different combinatorial optimization problems (for an overview see [12]).
Recently, a convergence proof for a generalized Ant System has been developed
by Gutjahr [13]. Generally, the Ant System algorithm consists of the iteration
of three steps:
Below we will discuss the first step in more detail. The second and third step
are reviewed only briefly (c.f. [4] for a more detailed description of these steps).
where d0i denotes the distance between the depot and customer i and Nu denotes
the set of unrouted customers. Further, k is the customer visited immediately
after customer j in the current solution, bik is the actual arrival time at customer
k, if i is inserted between customers j and k, while bk is the arrival time at
customer k before the insertion of customer i and Ri denotes the set of customers
assigned to the current tour after which customer i could feasibly be inserted.
Finally, typei is a binary indicator variable denoting whether customer i is a
linehaul (typei = 0) or a backhaul customer (typei = 1). The intuition is that
we want to be able to discriminate between linehaul and backhaul customers.
Note, that given ’inappropriate’ values for the parameters α, β, γ and δ the
attractiveness ηij can become zero. Note further, that our attractiveness function
is an extension of the function proposed by Solomon [14] for the VRPTW. More
precisely, our function reduces to Solomon’s function if δ is set to δ = 0.
Given the attractiveness we then compute the evaluation function of the
best insertion position for each customer i on the current tour by
τji +τik
κi = max {0, maxj∈Ri [ηij · 2·τjk ]} ∀i ∈ Nu ,
where τji denotes the pheromone concentration on the arc connecting locations
(customers or depot) j and i. The pheromone concentration τji contains infor-
mation about how good visiting two customers i and j immediately after each
other was in previous iterations. Note, that the same pheromone utilization is
done for route initialization, thus augmenting the attractiveness of initializing a
route with an unrouted customer i by the search-historic information.
Computing the attractiveness ηij reflects the tradeoff between detour and
delay costs associated with inserting a customer. This tradeoff is weighted by the
parameter β. A higher value for β puts higher emphasis on the detour and lower
emphasis on the delay. The tradeoff is also influenced by customer characteristics
such as distance to the depot or type and these characteristics are weighted
by the parameters α and γ, δ respectively. More precisely, higher α will favor
customers far from the depot, while higher δ will put more emphasis on the
customer type. Finally, γ influences the discrimination between linehaul and
backhaul customers.
Analyzing a Unified Ant System for the VRP and Some of Its Variants 305
4 Numerical Analysis
As testing our algorithm on all instances in the four classes would have been
computationally too expensive we performed our numerical analysis on selected
problem instances. First, to eliminate size effects we only considered problems
with approximately 100 customers. In fact, only the VRPB instances have 90
and 94 customers respectively as for this class 100 customer instances are not
available. Also, in order not to bias the results by taking ’easy’ or ’hard’ instances
we randomly chose 19 instances. These are:
– VRP (from [17]): vrp3, vrp8, vrp12, vrp14
– VRPTW (from [14]): r101n, r206n, c105n, c207n, rc104n, rc203n
– VRPB (from [18]): lhbh-i1, lhbh-i3, lhbh-j1, lhbh-j3
– VRPBTW (from [7]): bhr101b, bhr102a, bhr103b, bhr104c, bhr105c
L = M · F S + T T, (2)
where L denotes the total costs of a solution, F S denotes the fleet size found,
and T T corresponds to the total travel time (or distance). The parameter M
has to be chosen in a way to ensure that a solution that saves a vehicle always
outperforms a solution with a higher fleet size. More precisely we set M = 10000
in our experiments.
Finally, as we are interested in the interplay of the parameters of our insertion
algorithm with the characteristics of the different problems we analyzed the four
parameters of the insertion algorithm in the ranges α ∈ {0.5, 1, 1.5, 2}, β ∈
{0, 0.33, 0.66, 1}, γ ∈ {0, 0.33, 0.66, 1} and δ ∈ {0, 0.33, 0.66, 1} on the instances
described above.
This means that we tested 256 parameter constellations for each instance.
For each constellation we performed 1 run of 10 minutes. All our computations
were performed on a Pentium 3 with 900MHz. The code was implemented in C.
0,6
0,5
0,4
0,3
0,2
0,1
0
vrp3
vrp8
vrp12
vrp14
r101n
r206n
c105n
c207n
rc104n
rc203n
lhbh-i1
lhbh-i3
lhbh-j1
lhbh-j3
bhr101b
bhr102a
bhr103b
bhr104c
bhr105c
problem instances
Parameter Parameter
setting 0.5 1 1.5 2 setting 0 0.33 0.66 1
α 0.14% 0.07% 0.05% 0.03% β 0.03% 0.04% 0.04% 0.03%
γ 0.03% 0.03% 0.03% 0.03%
δ 0.03% 0.03% 0.03% 0.03%
Clearly, these results show that the parameter α has the strongest influence
on solution quality. The other three parameters show basically no sensitivity to
their actual setting. To understand this result let us consider the effects of the
parameters. Both the parameters α and δ are free parameters, i.e. they can take
any values. Moreover, higher levels for these parameters may implicitly increase
the number of positive evaluations κi . Thus, the number of customers on each
truck may increase which in turn leads to a possible reduction in fleet sizes. As
this is the main goal, higher values for α and δ should be expected to improve
the solution quality. However, both of the free parameters α and δ are at the
highest level we tested. This suggests that considering even higher values for
these parameters might further improve the performance of the algorithm. This
question is not considered in the remainder of this paper, but rather left for
future research.
Finally, the parameters β and γ have more indirect effects on the fleet size
such that their actual values will influence mainly the second objective namely
the total distance travelled.
against state of the art results. More precisely, we consider for each instance the
best known solution.1 Apart from the VRP instances - for those the standard
objective is to minimize total distance travelled only - all the other best known
results are based on the same lexicographic ordering of objectives we use.
In our comparison, we split the solutions into the fleet size and the total
distance travelled. Table 2 shows the results for all of the 19 instances.
Our results show that considering the first goal, namely the fleet size, our
algorithm improves the best known results in three cases, leading to a reduction
in fleet sizes equal to 1.46%. With respect to the second objective, our algorithm
deviates by 2.43% from the best known solution. Thus, the reduction in fleet
size is achieved through giving up some solution quality associated with the
total distance travelled. Overall, our algorithm being the first unified approach
for the four classes studied shows very promising behavior.
1
Note, that these best known results are taken from different sources as to date there
is no unified approach for the studied problem classes. Additional information about
the corresponding sources can be found at
http://www.bwl.univie.ac.at/bwl/prod/reimann/vrpresults.html.
Analyzing a Unified Ant System for the VRP and Some of Its Variants 309
References
1. Toth, P. and Vigo, D. (Eds.): The Vehicle Routing Problem. Siam Monographs on
Discrete Mathematics and Applications, Philadelphia (2002)
2. Cordeau, J. F., Gendreau, M., Laporte, G., Potvin, J. Y. and Semet, F.: A guide
to vehicle routing heuristics. Journal of the Operational Research Society 53 (5)
(2002) 512–522
3. http://www.top.sintef.no/
4. Reimann, M., Doerner, K., Hartl, R. F.: Insertion based Ants for Vehicle Routing
Problems with Backhauls and Time Windows. In: Dorigo, M. et al. (Eds.): Ant
Algorithms, Springer LNCS 2463, Berlin/Heidelberg (2002) 135–147
5. Bräysy, O. and Gendreau, M.: Metaheuristics for the Vehicle Routing Problem
with Time Windows. Sintef Technical Report STF42 A01025 (2001)
6. Toth, P. and Vigo, D.: VRP with Backhauls. In Toth, P. and Vigo, D. (Eds.):
The Vehicle Routing Problem. Siam Monographs on Discrete Mathematics and
Applications, Philadelphia (2002) 195–224
7. Gelinas, S., Desrochers, M., Desrosiers, J. and Solomon, M. M.: A new branching
strategy for time constrained routing problems with application to backhauling.
Annals of Operations Research. 61 (1995) 91–109
8. Duhamel, C., Potvin, J. Y. and Rousseau, J. M.: A Tabu Search Heuristic for
the Vehicle Routing Problem with Backhauls and Time Windows. Transportation
Science. 31 (1997) 49–59
9. Gendreau, M., Laporte, G., Musaraganyi, C. Taillard, E. D.: A tabu search heuris-
tic for the heterogeneous fleet vehicle routing problem. Computers and Operation
Research 26 (1999) 1153–1173
10. Cordeau, J. F., Gendreau, M., Laporte, G.: A Tabu Search Heuristic for Periodic
and Multi-Depot Vehicle Routing Problems. Networks 30 (1997) 105–119
11. Colorni, A., Dorigo, M. and Maniezzo, V.: Distributed Optimization by Ant
Colonies. In: Varela, F. and Bourgine, P. (Eds.): Proc. Europ. Conf. Artificial
Life. Elsevier, Amsterdam (1991) 134–142
12. Bonabeau, E., Dorigo, M. and Theraulaz, G.: Swarm Intelligence. Oxford Univer-
sity Press, New York (1999)
13. Gutjahr, W. J.: ACO algorithms with guaranteed convergence to the optimal so-
lution. Information Processing Letters. 82 (2002) 145–153
14. Solomon, M. M.: Algorithms for the Vehicle Routing and Scheduling Problems
with Time Window Constraints. Operations Research. 35 (1987) 254–265
15. Osman, I. H.: Metastrategy simulated annealing and tabu search algorithms for
the vehicle routing problem. Annals of Operations Research. 41 (1993) 421–451
16. Bullnheimer, B., Hartl, R. F. and Strauss, Ch.: A new rank based version of the ant
system: a computational study. Central European Journal of Operations Research
7(1) (1999) 25–38
17. Christofides, N., Mingozzi, A. and Toth, P.: The vehicle routing problem. In:
Christofides, N. et al. (Eds.): Combinatorial Optimization. Wiley, Chicester (1979)
18. Jacobs-Blecha, C., Goetschalckx, M.: The vehicle routing problem with backhauls:
properties and solution algorithms. Technical Report, School of Industrial and
Systems Engineering, Georgia Institute of Technology, Atlanta, Georgia (1992)
19. Jozefowiez, N., Semet, F. and Talbi, E.: Parallel and Hybrid Models for Multi-
objective Optimization: Application to the Vehicle Routing Problem. In Guervos,
J. M. et al. (Eds.): Parallel Problem Solving from Nature – PPSN VII, Springer
LNCS 2439, Berlin/Heidelberg (2002) 271–280