Swarm Intell (2007) 1: 135–151

DOI 10.1007/s11721-007-0005-x

Ant colony optimization for real-world vehicle routing

problems
From theory to applications

A.E. Rizzoli, · R. Montemanni · E. Lucibello ·

L.M. Gambardella

Received: 6 December 2006 / Accepted: 10 May 2007 / Published online: 13 September 2007
© Springer Science + Business Media, LLC 2007

Abstract Ant colony optimization (ACO) is a metaheuristic for combinatorial optimization

problems. In this paper we report on its successful application to the vehicle routing problem
(VRP). First, we introduce the VRP and some of its variants, such as the VRP with time
windows, the time dependent VRP, the VRP with pickup and delivery, and the dynamic
VRP. These variants have been formulated in order to bring the VRP closer to the kind of
situations encountered in the real-world.
Then, we introduce the basic principles of ant colony optimization, and we briefly present
its application to the solution of the VRP and of its variants.
Last, we discuss the applications of ACO to a number of real-world problems: a VRP
with time windows for a major supermarket chain in Switzerland; a VRP with pickup and
delivery for a leading distribution company in Italy; a time dependent VRP for freight distri-
bution in the city of Padua, Italy, where the travel times depend on the time of the day; and
an on-line VRP in the city of Lugano, Switzerland, where customers’ orders arrive during
the delivery process.

Keywords Ant colony optimization · Ant colony system · Vehicle routing problem ·
Dynamic VRP · Rich VRP · Real-world VRP

1 Introduction

The vehicle routing problem (VRP) concerns the transport of items between depots and
customers by means of a fleet of vehicles. Examples of VRPs are: milk delivery, mail de-
livery, school bus routing, solid waste collection, heating oil distribution, parcel pick-up
and delivery, dial-a-ride systems, and many others. Although finding the most cost efficient

A.E. Rizzoli, () · R. Montemanni · L.M. Gambardella

Istituto Dalle Molle di Studi sull’Intelligenza Artificiale (IDSIA), Galleria 2, 6928 Manno, Switzerland
e-mail: andrea@idsia.ch

E. Lucibello
AntOptima, via Fusoni 4, 6900 Lugano, Switzerland
136 Swarm Intell (2007) 1: 135–151

way to distribute goods across the logistic network is the main objective of supply-chain
systems, only in the early ’90s enterprise resource planning software vendors started to in-
tegrate tools to solve the VRP in supply chain management software (a review of software
for supply chain management can be found in Aksoy and Derbez 2003).
The practical interest of the VRP has spawn a number of studies, which tackled the prob-
lem from many sides. Yet, the VRP is combinatorially complex and, therefore, as the size
of the problem increases, it becomes harder and harder to obtain an exact solution for it in
a reasonable amount of time. Thus, even the most advanced exact solution methods impose
particular constraints on the problem instance, which are often violated when dealing with
real-world vehicle routing problems, leaving practitioners unsatisfied with the performance
and applicability of the algorithms.
Given the shortcomings of exact solution methods, researchers in the field of operations
research (OR) started to develop metaheuristics (Blum and Roli 2003), heuristic methods
that can be applied to a wide class of problems. One of the advantages metaheuristics have
over traditional optimization algorithms is their ability to produce a good suboptimal solu-
tion in short time. The integration of optimization algorithms based on metaheuristics, such
as tabu search (Glover and Laguna 1997), simulated annealing (Kirkpatrick et al. 1983), ant
colony optimization (Dorigo et al. 1996, 1999), and iterated local search (Lourenço et al.
2003), with advanced logistic systems for supply chain management opens new perspectives
for operations research applications in industry. In particular, for the solution of VRP and
its variants, a number of metaheuristics have been successfully applied, such as: simulated
annealing (Osman 1993), tabu search (Gendreau et al. 1994; Taillard et al. 1997), granular
tabu search (Toth and Vigo 2003), genetic algorithms (Van Breedam 1996), guided local
search (Kilby et al. 1999), variable neighborhood search (Bräysy 2003), greedy random-
ized adaptive search procedure (Resende and Ribeiro 2003), and ant colony optimization
(Gambardella 1999; Reimann et al. 2003).
In this paper we focus on ant colony optimization (ACO) (Dorigo and Stützle 2004), a
metaheuristic inspired by the foraging behavior of ant colonies. ACO has been used for the
approximate solution of a number of traditional OR problems, among which the job shop
scheduling problem, the quadratic assignment problem, the sequential ordering problem,
the graph coloring problem, and the shortest common supersequence problem (Dorigo and
Stützle 2004). More recently, ACO has been employed in a number of open shop scheduling
problems (Blum 2005), in optimal product design (Albritton and McMullen 2007), and has
also been used in some environmental problems, such as the design of a water distribution
network (Zecchin et al. 2007) or the planning of wells for groundwater quality monitor-
ing (Li and Chan Hilton 2007), thus proving its adaptability to very different domains of
application.
The flexibility of the ACO metaheuristic allowed its application to many vehicle rout-
ing problems where heterogeneous vehicle fleets, limitations on customer accessibility, time
windows, and the order imposed by pick-ups and deliveries considerably complicate the
problem formulation. These kinds of problems have been labeled as rich vehicle routing
problems (Hartl et al. 2006). Yet, real-world problems are even more complex; for instance,
travel times may be uncertain and may depend on traffic conditions, and not all customers’
orders may be perfectly known in time and dimension. These problem variants have been
called dynamic VRP and they are currently attracting a lot of research efforts, because of
their closeness to real-world traffic and distribution models (Zeimpekis et al. 2007). The
objective of this paper is to describe how ant colony optimization can be successfully used
to solve a number of VRP variants, both for some of the basic problem instances (the capac-
itated VRP, the VRP with time windows, the VRP with pickup and delivery) and for some
Swarm Intell (2007) 1: 135–151 137

of the dynamic extensions (the time dependent vehicle routing problem and the on-line ve-
hicle routing problem) where ACO has recently been applied and its ability to find efficient
solutions in a short time has been proven useful also in this setting.
The paper is structured as follows: first we outline the VRP with its static and dynamic
variants; then we introduce ACO, providing an overview of the main ACO algorithms; fi-
nally we discuss the application of ACO to the VRP. In this latter section, devoted to appli-
cations, we first introduce two large-scale industrial applications, showing how ACO can be
successfully applied in the day-to-day operations of large real-world distribution processes,
then we present the application of ACO to dynamic VRPs, showing its applicability in the
context of urban freight distribution.

2 Vehicle routing problems

Finding optimal routes for a fleet of vehicles performing assigned tasks on a number of spa-
tially distributed customers can be formulated as a combinatorial optimization problem: the
vehicle routing problem. A solution of this problem is the best route serving all customers
using a fleet of vehicles, respecting all operational constraints, such as vehicle capacity and
the driver’s maximum working time, and minimizing the total transportation cost.
The algorithm solving the problem requires the definition of an objective function that
may include multiple objectives that are often conflicting. The most common objective is
the minimization of transportation costs as a function of the traveled distance or of the travel
time; the number of vehicles can be minimized expressing the costs associated with vehi-
cles and drivers. Vehicle efficiency, expressed as the percentage of load capacity weighted
by distance, can be taken into account. “Soft” constraints, which can be violated paying a
penalty, can be included. For instance, if a customer is not served according to the agreed
time schedule, a penalty might have to be paid. Road pricing schemes can also be consid-
ered, for example, attributing a higher cost to routes through city centers.
The elements that define and constrain each model of the VRP are: the road network,
describing the connectivity among customers and depots; the vehicles, transporting goods
between customers and depots on the road network; the customers, which place orders and
receive goods.
The road network graph can be obtained from a detailed map of the distribution area
on which the depots and the customers are geo-referenced. Standard algorithms can then
be used to find the shortest routes between all couples of nodes in order to build the travel
cost matrix. The matrix coefficients can represent the time required to travel from node i to
node j , or the distance between the nodes, or any other metric that measures the travel cost.
According to the adopted metric, different instances of the VRP may arise. For instance,
if the travel time on edges depends on the time of the day, then we encounter the time
dependent VRP (see Sect. 4.3.1 below).
The fleet of vehicles and their characteristics also impose constraints on the vehicle rout-
ing model. The fleet can be homogeneous, if all vehicles are equal in their characteristics,
heterogeneous if this is not the case. Most real-world fleets are heterogeneous. Mechanical
features (length, weight, width) and configuration (trailer, semi-trailer, van, etc.) constrain
the ability of a vehicle to access road segments. For instance, a vehicle cannot travel on
some arcs of a road network because of excessive weight or dimensions. On-board equip-
ment, such as loading/unloading devices, may also impose constraints that depend on the
type of customer to be served. Capacity constraints, stating the maximum load to be trans-
ported by a vehicle, are also important.
138 Swarm Intell (2007) 1: 135–151

Each customer requests a given amount of goods, an order, which must be delivered
or collected at the customer location. Time windows within which the customer must be
served can be specified. These time windows can be single (only one continuous interval)
or multiple (disjoint intervals). If time windows cannot be violated at any cost they are said
to be “hard”; on the other hand, when a penalty is paid in case of violation, time windows
are said to be “soft”. Finally, the vehicle routing model can also include an estimation of the
loading and unloading times at the customer—the so-called service time.

2.1 Basic problems of the vehicle routing class

Combining the various elements of the problem, we can define a whole family of different
VRPs. Toth and Vigo (2001b) present a detailed overview of the various VRPs.
The capacitated vehicle routing problem (CVRP) is the basic version of the VRP. The
name derives from the constraint of having vehicles with limited capacity. Customer de-
mands are deterministic and known in advance. Deliveries cannot be split, that is, an order
cannot be served using two or more vehicles. The vehicle fleet is homogeneous and there is
only one depot. The objective is to minimize the total travel cost, usually expressed as the
traveled distance required to serve all customers. The CVRP is NP-hard (Labbé et al. 1991)
and the size of the problems which can be solved exactly in a reasonable time is up to 50
customers, using the branch-and-bound, branch-and-cut, and set-covering approaches (see
Toth and Vigo 2001a).
When constraints on the delivery times are present, we have a vehicle routing problem
with time windows (VRPTW): the capacity constraint still holds and each customer i is as-
sociated with a time window [ai , bi ] and with a service time si . VRPTW is also NP-hard, and
even finding a feasible solution to the VRPTW is an NP-hard problem (Savelsbergh 1985).
Good overviews on the VRPTW formulation and on exact, heuristic, and metaheuristic ap-
proaches to its solution can be found in Mester and Bräysy (2005), Li et al. (2005), and in
Kytöjoki et al. (2007). Kallehauge et al. (2006) have solved VRPTW problems with 400 and
1000 customers by Lagrangian relaxation, but the problem formulation requires hard time
windows.
In the VRP with pick-up and delivery (VRPPD) the transport items are not originally
concentrated in the depots, but they are distributed over the nodes of the road network.
A transportation request consists in transferring the demand from the pick-up point to the
delivery point. These problems always include time windows for pick-up and/or delivery.
A review of various approaches to the solution of the VRPPD is presented in (Desaulniers
et al. 2000).
In all of these approaches the problem data is supposed not to change, neither during
the planning phase (computation of the solution), nor during the management phase (im-
plementation of the solution). More realistic situations might require the relaxation of this
assumption.

2.2 Dynamic extensions of the VRP

The static formulations of the VRP, where customer demand is deterministic and travel
times do not depend on the time of the day, have proven to be successful in modeling many
practical problems. This is especially true for the VRPTW and VRPPD extensions. Yet, the
availability of online information on the traffic conditions and the possibility of monitoring
the vehicles’ positions via the global positioning system, together with the online update
of customer orders, can considerably change the problem settings. The availability of this
Swarm Intell (2007) 1: 135–151 139

information has a price: the assumption of time-invariancy must be relaxed and data become
time-dependent. Moreover, using data on current traffic conditions to estimate travel times
requires the relaxation of the assumption of determinism, introducing uncertainty and adding
another level of complexity to the problem.
In the literature these problems have been labeled with different terms: probabilistic,
dynamic, and stochastic vehicle routing. These terms are often interchanged, but in practice
we can assume that, in general, we are confronted with a dynamic vehicle routing problem,
where problem data are often generated by a stochastic process. For instance, a stochastic
process can be assumed to be responsible for the presence or absence of the customers,
for the quantity of their orders, and for the travel and service times (Laporte and Louveaux
1998).
Stochastic customers and demands are typically formed when the planning horizon is
longer than the horizon of the currently available data. These kinds of problems have been
extensively studied (see, for instance, Gendreau et al. 1996); Bianchi et al. (2004) also stud-
ied how various metaheuristics can used in the solution of the VRP with stochastic demands.
One approach to solve the problem of unknown orders is to remove uncertainty by
processing new orders as they come, in batches of variable size, according to what is called a
reactive strategy. Potvin et al. (2006) list a number of reactive dynamic strategies for vehicle
routing and scheduling problems; in the case of VRP this has been called the on-line vari-
ant (OLVRP). Among possible applications of the OLVRP we find feeder systems, which
typically are local dial-a-ride systems aimed at feeding another, wider area, transportation
system at a particular transfer location (Gendreau and Potvin 1998; Psaraftis 1988, 1995).
Another side of the problem is the presence of stochastic travel and service times, which
are very frequent in urban environments. Especially with respect to travel times, the variabil-
ity can be very high and considerably affect the solution. In the time dependent VRP variant
(TDVRP) the variability can be reduced if one assumes that travel times are nearly con-
stant within time periods in a day. This is quite true for peak and off-peak traffic conditions,
which are observed in most cities. Ichoua et al. (2003) present a structured introduction to
the problem and a model formulation.

3 The ACO metaheuristic and an overview of ACO algorithms

Ant colony optimization (Dorigo et al. 1999) is an optimization framework inspired by the
observation, made by ethologists, that ants use pheromone trails to communicate informa-
tion regarding shortest paths to food. A moving ant lays some pheromone (in varying quanti-
ties) on the ground, thus marking a path with a trail of this substance. An isolated ant moves
mostly randomly and when it detects a previously laid pheromone trail it can decide, with
high probability, to follow it, thus reinforcing the trail with its own pheromone. The collec-
tive behavior that results is a form of autocatalytic behavior where the more ants follow a
trail, the more attractive for other ants it becomes. The process is thus characterized by a
positive feedback loop, where the probability with which an ant chooses a path increases
with the number of ants that previously chose the same path.
The above process inspired the ACO metaheuristic. The main elements are artificial ants
(from now on simply ants), simple computational agents that individually and iteratively
construct solutions for the problem, which has been modeled as a graph. Ants explore the
graph visiting nodes connected by edges. A problem solution is an ordered sequence of
nodes. The search process is executed in parallel over several constructive computational
threads. A dynamic memory structure, inspired by the pheromone laying process, which
140 Swarm Intell (2007) 1: 135–151

incorporates information on the effectiveness of previously obtained results, guides the con-
struction process of each thread. Intermediate partial problem solutions are seen as states;
(j )
at each iteration k of the algorithm each ant moves from state xk(i) to xk+1 , expanding the
partial solution from node i adding node j .
Ant system (AS) (Dorigo et al. 1996) was the first ACO algorithm to be proposed. It is
organized in two main stages: construction of a solution, and update of the pheromone trail.
In AS each ant builds a solution. An ant is in a given state and it computes a set of feasible
expansions from it. The ant selects the move to expand the state taking into account the
following two values:
1. the attractiveness ηij of the move, as computed by some heuristic indicating the a priori
desirability of that move;
2. the pheromone trail level τij of the move, indicating how useful it has been in the past
to make that particular move; it represents, therefore, an a posteriori indication of the
desirability of that move.
Given the attractiveness and the pheromone trail level, an ant chooses to visit node j
from node i according to the following probability:
α β
⎧
⎨  [τij ] [ηij ]
, if j ∈ Ω,
pij = [τ ] [ηih ]β
α (1)
⎩ h∈Ω ih
0, otherwise,

where Ω is the set of nodes which can be visited starting from i. The parameters α and β
weigh the influence of trails and attractiveness.
Once a solution is obtained, pheromone trails are updated. First, the pheromone is evap-
orated on all arcs, in order to progressively forget bad solutions; then all ants deposit
pheromone on the arcs which are part of the solutions they have just computed.
Dorigo and Stützle (2004) distinguish between variants and extensions of AS. The vari-
ants differ from the original algorithm mainly in the pheromone update rule. Among variants
we find: elitist ant system (Dorigo 1992); rank-based ant system (Bullnheimer et al. 1999);
and MAX -MIN ant system (Stützle and Hoos 2000), which uses bounding techniques
to limit the possible values for pheromone on arcs. Extensions display more substantial
changes in the algorithm structure. Among the extensions we find: approximate nondeter-
ministic tree search (ANTS, Maniezzo and Carbonaro 2000) exploiting the use of lower
bounds in the computation of a solution; D-ants (Reimann et al. 2002, 2004), which makes
use of the savings algorithm; the hyper-cube framework for ACO (Blum and Dorigo 2004),
which rescales pheromone values between 0 and 1; beam-ACO (Blum 2005), which hy-
bridizes ACO with beam search; and, finally, ant colony system (ACS, Dorigo and Gam-
bardella 1997) on which we focus in the remainder. In brief, ACS differs from AS for a
revised rule used in the tour construction algorithm, and for the use of both local and global
updates of the pheromone trails.
During tour construction ACS uses the pseudo-random-proportional transition rule. Let
i be the last node visited in the partial solution under construction. If q is a random variable
uniformly distributed over [0, 1], with probability q (exploitation) the next chosen node j is
the one maximizing the product τij · ηij , while with probability 1 − q (exploration) the node
j is chosen probabilistically via Monte Carlo sampling using (1). Notice that the pseudo-
random-proportional rule provides a straightforward way to balance between exploration of
new states and exploitation of a priori and accumulated knowledge. Parameter q, that is, the
probability of carrying out exploitation rather than exploration, is usually set to 0.9.
Swarm Intell (2007) 1: 135–151 141

ACS uses two different types of pheromone updates: local and global. Local update
is performed every time an ant decides to use the edge from i to j in its solution; the
pheromone τij is then modified:

τij = (1 − ρ) · τij + ρ · τ0 , (2)

where τ0 is the initial pheromone value defined as τ0 = 1/(n · Jnn ), where n is the number
of customers in the solution and Jnn the objective function produced by the execution of
one ACS iteration without the pheromone component (this is equivalent to a probabilistic
nearest neighbor heuristic).
Global update is carried out at the end of each iteration, when each ant has constructed
a complete solution. The only edges modified are those belonging to the best solution gen-
erated so far, ψ ∗ . The objective function evaluates ψ ∗ to the value J ∗ . The global update
formula is:
1
τi ∗ j ∗ = (1 − ρ) · τi ∗ j ∗ + ρ · ∗ , ∀ i ∗ , j ∗ ∈ ψ ∗ .
 
(3)
J
ACS has been shown to be very efficient in solving routing problems. In the next sections
we show how it has been applied to the solution of some problem instances of the VRP
class, ranging from the static case (VRP with time windows, and VRP with time windows
and pick-up and delivery constraints) to the dynamic case (time dependent VRP and on-line
VRP).

4 Solving the VRP with ACO

Sales and distribution processes require the ability to forecast customer demand and to opti-
mally plan the distribution of the products to the consumers. These two strategic activities,
forecast and optimization, must be tightly interconnected in order to improve the perfor-
mance of the system as a whole (Gambardella et al. 1998).
In Fig. 1 the workflow of a distribution-centered company is sketched. The sales depart-
ment generates new orders by contacting the customers (old and new) to check whether
they need a new delivery. The effectiveness of this operation can be increased thanks to
inventory management modules, which estimate the demand of every customer, indicating
the best re-order time for each of them. New orders are then processed by the planning de-
partment, which, according to the quantities requested, the location of the customers, and
the time windows for the delivery, decides how many vehicles to employ and computes the
best routes for the delivery, in order to minimize the total travel time and space. This task is
assisted by a vehicle routing algorithm, represented by the OPTIMIZE block. The vehicle
tours are then assigned to the fleet, which is monitored by the fleet operational control sta-
tion, which monitors the evolution of deliveries in real time. This process is assisted by the
SIMULATE/MONITOR/RE-PLAN module, which allows re-planning online in face of new
urgent orders, which were not available during the previous off-line planning phase. Finally,
after vehicles have returned to the depot, delivery data are off-loaded and transferred back
to the company database.
In the next sub-sections, we describe a number of real-world applications, where ACO
has been used for the implementation of the vehicle routing algorithms that are executed in
the OPTIMIZE block of Fig. 1.
142 Swarm Intell (2007) 1: 135–151

Fig. 1 The forecast–optimize–simulate loop: the role of optimization in the efficient management of a dis-
tribution process

4.1 A VRPTW application for the secondary level in the supply chain

In this application the client is one of the major supermarket chains in Switzerland. The
problem is to distribute palletized goods to more than 600 stores, all over Switzerland. The
stores order daily quantities of goods to replenish their local stocks. They want the goods to
be delivered within time windows, in order to plan in advance the daily availability of their
personnel, allocating a fraction of their time to inventory management tasks.
There are three types of vehicles: trucks (capacity: 17 pallets), trucks with trailers (35
pallets), and tractor units with semi-trailers (33 pallets). Whether a vehicle can access a
store or not depends on the store location. In some cases the truck with trailer can leave the
trailer at a previous store and then continue to other less accessible locations. The number of
vehicles is assumed to be infinite, since transport services can be purchased on the market
according to the needs.
The road network graph has been computed using digital road maps. The distances in
kilometers between couples of stores have been rescaled using a speed model, which de-
pends on the distance: longer distances allow a higher average speed. For instance, if the
distance is less than 5 km, the average speed is 20 km/h; if the distance is more than 90 km,
the speed is 60 km/h; in between there is a range of other speed values. The data have been
collected over many years and they have been validated by the drivers’ experience. The time
to set-up a vehicle for unloading and the time required to hook/unhook a trailer are constant.
The service time is variable and depends on the number of pallets to unload.
All the routes must be performed in one day, and the client imposes an extra constraint
stating that a vehicle must perform its latest delivery as far as possible from the inventory,
since it could be used to perform extra services on its way back. These extra services were
not included in the planning by explicit request of the client.
Swarm Intell (2007) 1: 135–151 143

4.1.1 The algorithm

This problem was modeled as a VRPTW, and solved by an implementation of the MACS-
VRPTW algorithm (Gambardella 1999), named ANTROUTE. MACS-VRPTW is the most
efficient ACO algorithm for the VRPTW and one of the most efficient metaheuristics overall
for this problem. ANTROUTE adds to MACS-VRPTW the ability to handle the choice of
the vehicle type: at the start of each tour the ant chooses a vehicle. A waiting cost was also
introduced in order to prevent vehicles arriving too early at the stores.
The central idea of the MACS-VRPTW algorithm is to use two ant colonies (MACS
stands for multi ant colony system). One colony, named ACS-VEI, minimizes the vehicles
while the other one, named ACS-TIME, minimizes time. The two colonies are completely
independent, since each one has its own pheromone trail, but they collaborate by sharing the
variable ψ ∗ , which describes the best solution found so far. Each colony is composed of a
number of ants. Every ant in the colony tries to build a feasible solution to the problem.
The algorithm works as follows: first compute a feasible initial solution ψ ∗ , with a num-
ber of vehicles v, using a nearest neighbor heuristic. ACS-VEI is then started: it tries to
find a feasible solution ψ ∗ using v − 1 vehicles. ACS-TIME is also started: given v vehi-
cles, it tries to minimize the total time required to serve all the customers. When ACS-VEI
finds a feasible solution with v − 1 vehicles, ψ ∗ is updated and ACS-TIME is restarted with
v = v − 1. ACS-TIME serves the purpose of refining the solutions obtained by ACS-VEI,
which has no feeling for the travel time, since its objective function is independent of it.
The constructive step of both colonies is based on a procedure that, starting from the
current node i, computes the set of all feasible nodes. These are the nodes j still to be
visited and such that the time of arrival at node j and the load are compatible with the time
window and the delivery of the quantity qj of goods requested by node j . The probabilistic
choice is made according to the pseudo-random-proportional rule described in Sect. 3. In
order to account for the effect of time windows, distance becomes a function of the time
window: a node can appear “closer” if the end of its time windows is near. The end of the
construction of a solution by all ants in the colony marks the end of one algorithm iteration.
The pheromone trails are updated both locally and globally according to the ACS algorithm.
Note that in the ACS-VEI colony the ants usually construct infeasible solutions, that is,
not all customers can be visited under the constraints given.
Given that the algorithm uses two colonies, at the end of each cycle, pheromone trails
are globally updated for two different solutions: ψ ACS-VEI , the infeasible solution with the
highest number of visited customers, and ψ ∗ , the feasible solution with the lowest number
of vehicles and the shortest travel time. Thus, pheromone is updated also on arcs that are not
included in a feasible solution, which could still become feasible in the next iteration.
ACS-TIME uses a local search procedure to improve the quality of the feasible solutions,
which is similar to the CROSS procedure (Taillard et al. 1997). Given that the CROSS
procedure requires a feasible solution to operate on, such a local search can not be applied
during ACS-VEI, where unfeasible solutions can still be returned at the end of a cycle. Both
ACS-TIME and ACS-VEI try to repair infeasible solutions by inserting unvisited customers.

4.1.2 Results

The first tours computed by ANTROUTE were not accepted as feasible by the human tour
planners, even if the performance was considerably higher than theirs and no explicit con-
straints were violated. Thus, a further modeling iteration was required, to let “hidden” con-
straints emerge. The human planners were actually using a regional planning strategy, that
144 Swarm Intell (2007) 1: 135–151

Table 1 Comparison of the

computer-generated vs. Human AR-RegTW AR-Free
man-made tours in the VRPTW planner
application
Total number of tours 2056 1807 1614
Total km 147271 143983 126258
Average truck loading 76.91% 87.35% 97.81%

Fig. 2 The relationship between

the time window width and the
number of vehicle routes

led to petal shaped tours. This preference was included in the reformulation of the problem,
but at the same time we tried to loosen the constraint a bit. We attributed stores to distribu-
tion regions, but at the same time we allowed stores near the border of the distribution region
to also belong to the neighboring region. This allowed the generation of tours which were
slightly worse than the unconstrained solution, but, nevertheless, better than the solutions
found by the human planners. In Table 1 we present the results obtained by ANTROUTE
compared with those of the human planners. ANTROUTE was run under two configurations:
AR-RegTW, with regional planning and 1-hour time windows; AR-Free, where the regional
and the time windows constraints were relaxed. The problem was to distribute 52000 pal-
lets to 6800 customers over a period of 20 days. Every day ANTROUTE was run on the
available set of orders and it took about 5 minutes to find a solution. At the same time, the
planners were at work and it took them at least 3 hours to find a solution. At the end of the
testing period, the performances of the algorithm and of the planners have been compared
using the same objective function.
The advantage of an algorithm able to find the solution to an otherwise very hard problem
in such a short time is the possibility of using it as a strategic planning tool. In Fig. 2 it is
shown how running the algorithm with wider time-windows at the stores returns a smaller
number of tours, which can be translated in a substantial reduction of transportation costs.
The logistic manager can, therefore, use the optimization algorithm as a tool to investigate
how to re-design the time-windows in the stores.

4.2 A VRPPD application for the primary level in the supply chain

In this application the client is a major logistics operator in Italy. The distribution process
involves moving palletized goods from factories to inventory stores, before they are finally
distributed to shops. A customer in this vehicle routing problem is either a pick-up or a
delivery point. There is no central depot, and approximately 1000–1500 trucks per day are
used. Routes can be performed within the same day, over two days, or over three days, since
Swarm Intell (2007) 1: 135–151 145

the Italian peninsula is quite long and there’s a strict constraint on the maximum number of
hours per day that a driver can travel. All pick-ups of a tour must take place before deliveries.
Orders cannot be split among tours. Time windows are associated with each store.
There is only one type of truck: tractor with semi trailer. The load is measured in pal-
lets, in kilograms, and in cubic meters. There are capacity constraints on each one of these
measurement units, and the first one that is exceeded triggers the violation of the constraint.
The availability of vehicles is assumed to be infinite, since they are provided by flexible sub-
contractors. Sub-contractors are distributed all over Italy and, therefore, vehicles can start
their routes from the first assigned customer, and no cost is incurred in traveling to the first
customer in the route.
The road network graph has been elicited from digital road maps, computing the shortest
path between each couple of stores. The travel times are computed according to the traveled
distance, given the average speed that can be sustained on each road segment according to its
type (highway, extraurban road, urban road). Loading and unloading times are assumed to
be constant. This is a rough approximation imposed by the client, since he has been unable
so far to provide better estimates. The client also imposed another constraint, related to the
same problem, setting a maximum number of cities to visit per tour (usually less than six).
Note that more than one customer can reside in a city. Moreover, the client requested that
the distance between successive deliveries should be limited by a parameter.

4.2.1 The algorithm

The problem can be modeled as a VRP with pickup and delivery and time windows  (VRP-
PDTW). The objective function measures the average tour efficiency, f = N i=1 ei /N ,
where ei is the efficiency of tour i: the occupancy ratio of a vehicle
Mi over the traveled distance
within the tour. It is computed according to the formula ei = j =1 qj lj /(Qi Li ), where Mi
is the number of orders in the ith tour; qj is the number of pallets in the j th order; lj is the
distance between source and destination points of the j th order; Qi is the capacity of the
vehicle serving the ith tour, and Li is the total length of the ith tour.
The ANTROUTE algorithm has also been used in this context, but since there is a single
objective—to maximize average efficiency—it has been adapted removing the ant colony
minimizing the number of vehicles. The first step of an artificial ant is to select the starting
city. Since this is a pickup and delivery problem, each source node must be paired with the
corresponding destination node, and the search space is, therefore, harder to explore than in
a delivery problem. The algorithm tries to simplify exploration using an approximation of
the delivery phase, assuming that all deliveries will be performed in the reverse order with
respect to pickups. Thus, a first stage local search exchanges nodes between tours, while
preserving the order of deliveries; later, another local search procedure is applied, in which
nodes are exchanged within the same tour.

4.2.2 Results

Table 2 summarizes the comparison between man-made and computer-generated tours over
a testing period of two weeks. A noticeable improvement in the efficiency of computer-
generated tours can be observed.
It is also interesting to remark how the algorithm performance is correlated with the dif-
ficulty of the problem, which is related to the number of orders to satisfy. In Fig. 3 we plot
on the x-axis the efficiency of the man-made tours, and on the y-axis the efficiency improve-
ment obtained using the computer-generated tours. When the problem is easy, because it
146 Swarm Intell (2007) 1: 135–151

Table 2 Comparison of the computer-generated vs. man-made tours in the VRPPDTW application

Human ANTROUTE Absolute Relative

planner difference difference

Total nr of tours 471.5 460.8 −10.7 −2.63%

Total km 175441 173623 −1818.2 −1.32%
Efficiency 84.08% 88.27% + 4.19% –

Fig. 3 Comparing man-made

and computer-generated tours.
Higher efficiency improvements
are observed when the human
planner performance is lower.
The dots are experimental values,
and the solid line is a regression
on those values

involves a limited number of orders, and the human planner performs well, the computer is
not able to provide a significative improvement, but when the planner starts to fail coping
with the problem complexity, and the performance decreases, the gain in using the algorithm
sensibly increases.

4.3 Time dependent VRPTW in the city of Padua

The city of Padua, Italy, set up a logistic platform to collect all incoming goods to be distrib-
uted to a number of shops in the city center. Such a platform aims at a better organization of
the flow of goods into the city center, which is affected by traffic congestion problems and
where loading and unloading space is scarcely available. Centralized planning of vehicle
routes can sensibly reduce pollution and traffic problems due to commercial transport. For
this purpose, Donati et al. (2007) have developed an algorithm solving the time dependent
VRPTW for a logistic platform serving the city of Padua. In the study, the central depot was
open from 8 am to 6 pm and traffic data on the Padua road network during that period was
collected. Four time intervals, with similar traffic patterns and the relative travel speeds on
network arcs, were identified. A set of 30 customers was considered.

4.3.1 The algorithm

The basic idea in the ACO algorithm of Donati et al. is to define a pheromone trail that is
time dependent assuming that the travel times over the arcs of the graph depend on the time
of the day. While the variation of travel times over time is continuous, it can be assumed
that there are some distinctive time slices during one day when they are roughly constant.
It is assumed that the duration of one working day can be partitioned in l time slices and,
therefore, the pheromone trails can be described by τij (l), with l ∈ Tl , where Tl is the set
Swarm Intell (2007) 1: 135–151 147

of time slices into which the working time horizon is split. The objective is to minimize the
total travel time.
The algorithm, based on MACS-VRPTW (see Sect. 4.1.1), then builds a solution making
a probabilistic choice to select the next node j starting from i using the standard equation
(1). The attractiveness ηij of the next node is given by:

1
ηij (t) = , (4)
fij (t) + wj

where fij (t) is the travel time from i to j evaluated at time t ∈ Tl and wj is the waiting time
at node j .
Pheromone updating is carried out as described in Sect. 3, independently for the
pheromone corresponding to each time slice.

4.3.2 Results

The authors compared the solution of the VRPTW using the time dependent variant with a
solution of the same problem where the travel times on the road arcs were constant, depend-
ing only on the distance. In a series of nine tests, where customers were chosen randomly
out of a set of real customers, it turned out that the time dependent variant performed 7%
better than the standard VRPTW algorithm.

4.4 On-line VRP for fuel distribution

A leading Swiss fuel oil distribution company, which serves its customers from its main
depot located near Lugano with a fleet of 10 vehicles, observed that during every Winter
season there was always a subset of their customers that ran out of fuel and had to place
urgent orders. These unexpected orders affects the planned delivery routes of the vehicles,
and the vehicle routing problem becomes very “dynamic”, since a noticeable percentage of
orders must be fulfilled after the vehicles have already left the depot.
The objective of this study was to evaluate the impact of a reactive strategy for vehicle
routing, starting from an analysis of the data collected in periods when urgent deliveries
were in high demand. From the company data base, a sample of 50 customers was randomly
selected, and travel times among them were calculated. In the company records, customers
randomly appeared during the working day with random requests for a quantity of fuel to be
delivered. A working day of 8 hours was considered, assuming a service time of 10 minutes
for each customer. The cut-off time, after which the new orders received were postponed to
the following working day, was set to 4 hours.

4.4.1 The algorithm

The problem description above matches the on-line VRP variant, where new orders can
be assigned to vehicles which have already left the depot (e.g., parcel collection, feeder
systems, fuel distribution, etc.).
To solve the on-line fuel oil distribution problem, Montemanni et al. (2005) have devel-
oped an ACO-inspired algorithm, ACS-DVRP, based on the decomposition of the on-line
VRP into a sequence of static VRPs. There are three main elements in the algorithm ar-
chitecture: the event manager, the ant colony algorithm, and the pheromone conservation
strategy.
148 Swarm Intell (2007) 1: 135–151

Table 3 Experimental results on

the case study of Lugano nts 200 100 50 25 10 5
tacs 144 288 576 1152 2880 5760
tls 15 30 60 120 240 480

Travel time 12702 12422 10399 9744 10733 11201

The event manager receives new orders and keeps track of the already served orders
and of the position and the residual capacity of each vehicle. This information is used to
construct the sequence of static VRP-like instances. The working day is divided into time
slices and for each of them a static VRP, which considers all the already received (but not
yet executed) orders, is created. New orders received during a time slice are postponed until
its end. At the end of each time slice, customers whose service time starts in the next time
slice (according to the solution of the last static VRP) are assigned to the vehicles. They will
not be taken into account in the following static VRPs.
The ACS algorithm employed is based on the one described in Sect. 4.1.1 for the
VRPTW. The single ant colony is in charge of minimizing the total travel time.
Finally, the pheromone conservation strategy is based on the fact that, once a time slice
is over and the relative static problem has been solved, the pheromone matrix contains infor-
mation about good solutions. As each static problem is potentially very similar to the next
one, this information is passed on to the next problem (Guntsch and Middendorf 2001): if a
couple of customers appears in both the previous and the current time slice, the pheromone
on the arcs connecting two nodes is brought forward as a fraction of its value in the previous
problem.

4.4.2 Results

Algorithm ACS-DVRP was executed on a number of test problems, obtained varying the
number nts of time slices into which the working day was divided. As the size of each prob-
lem in a time slice increases as the length of the time slice decreases, the time tacs allocated
to executing the ant colony system and the time tls dedicated to local search improving the
solution were adjusted accordingly. In particular, the ratio between tacs and tls was kept ap-
proximately equal to 10.
The first three rows of Table 3 define the settings of the experiments, that is, the values of
parameters nts , tacs , and tls . The fourth row shows the total travel time of the solutions found
by the ACS-DVRP algorithm.
The results suggest that, for the case study analyzed, good values for nts are between
10 and 50. In particular, 25 seems to be the best choice. Large values of nts did not lead
to satisfactory results because optimization was restarted too often, before a good local
minimum could be reached. On the other hand, when nts was too small, the system was
not able to take advantage of information on new incoming orders.

4.5 Conclusions

In this paper we have described how the ant colony optimization metaheuristic can be suc-
cessfully used to solve a number of variants of the basic vehicle routing problem. We pre-
sented two industrial-scale applications of ACO for the solution of static VRP problems:
a VRP with time windows and a VRP with pickup and delivery. We then focused our at-
tention on two important dynamic variants of the VRP: the time dependent VRP, and the
Swarm Intell (2007) 1: 135–151 149

on-line VRP. Both these problems are receiving increasing attention due to their relevance
to real world problems, in particular, for distribution in urban environments.
In conclusion, after more than ten years of research, ACO has been shown to be one of
the most successful metaheuristics for the VRP and its application to real-world problems
demonstrates that it has now become a fundamental tool in applied operations research.

Acknowledgements The authors wish to acknowledge the contribution by F. Oliverio, of AntOptima, in

the development of the ANTROUTE code. They also wish to thank the anonymous reviewers and the editors
who greatly helped in improving the focus and organization of this paper.

References

Aksoy, Y., & Derbez, A. (2003). Software survey: supply chain management. OR/MS Today, 30(3), 1–13.
Albritton, M. D., & McMullen, P. R. (2007). Optimal product design using a colony of virtual ants. European
Journal of Operational Research, 176(1), 498–520.
Bianchi, L., Birattari, M., Chiarandini, M., Manfrin, M., Mastrolilli, M., Paquete, L., Rossi-Doria, O., &
Schiavinotto, T. (2004). Metaheuristics for the vehicle routing problem with stochastic demands. X. Yao,
et al. (Eds.), Lecture notes in computer science: Vol. 3242. Parallel problem solving from nature—PPSN
VIII (pp. 450–460). Berlin: Springer.
Blum, C. (2005). Beam-ACO—Hybridizing ant colony optimization with beam search: an application to open
shop scheduling. Computers and Operations Research, 32(6), 1565–1591.
Blum, C., & Dorigo, M. (2004). The hyper-cube framework for ant colony optimization. IEEE Transactions
on Systems, Man and Cybernetics—Part B: Cybernetics, 34(2), 1161–1172.
Blum, C., & Roli, A. (2003). Metaheuristics in combinatorial optimization: overview and conceptual com-
parison. ACM Computing Surveys, 35(3), 268–308.
Bräysy, O. (2003). A reactive variable neighborhood search for the vehicle routing problem with time win-
dows. INFORMS Journal on Computing, 15(4), 347–368.
Bullnheimer, B., Hartl, R. F., & Strauss, C. (1999). A new rank-based version of the ant system: a computa-
tional study. Central European Journal of Operations Research, 7(1), 25–38.
Desaulniers, G., Desrosiers, J., Erdmann, A., Solomon, M. M., & Soumis, F. (2000). VRP with pickup and
delivery. In P. Toth & D. Vigo (Eds.), The vehicle routing problem (pp. 225–242). Philadelphia: SIAM.
Donati, A. V., Montemanni, R., Casagrande, N., Rizzoli, A. E., & Gambardella, L. M. (2007, in press). Time
dependent vehicle routing problem with a multi ant colony system. European Journal of Operational
Research.
Dorigo, M. (1992). Optimization, learning and natural algorithms. PhD thesis, Dipartimento di Elettronica,
Politecnico di Milano, Italy.
Dorigo, M., & Gambardella, L. M. (1997). Ant colony system: A cooperative learning approach to the trav-
eling salesman problem. IEEE Transactions on Evolutionary Computation, 1(1), 53–66.
Dorigo, M., & Stützle, T. (2004). Ant colony optimization. Cambridge: MIT Press.
Dorigo, M., Maniezzo, V., & Colorni, A. (1996). Ant system: optimization by a colony of cooperating agents.
IEEE Transactions on Systems, Man and Cybernetics—Part B: Cybernetics, 26(1), 29–41.
Dorigo, M., Di Caro, G., & Gambardella, L. M. (1999). Ant algorithms for discrete optimization. Artificial
Life, 5, 137–172.
Gambardella, L. M., Rizzoli, A. E., & Zaffalon, M. (1998). Simulation and planning of an intermodal con-
tainer terminal. Simulation, 71(2), 107–116.
Gambardella, L. M., Taillard, É., & Agazzi, G. (1999). MACS-VRPTW: a multiple ant colony system for
vehicle routing problems with time windows. In D. Corne, M. Dorigo & F. Glover (Eds.), New ideas in
optimization (pp. 63–76). London: McGraw–Hill.
Gendreau, M., & Potvin, J.-Y. (1998). Dynamic vehicle routing and dispatching. In T. G. Crainic & G. Laporte
(Eds.), Fleet management and logistic (pp. 115–226). Berlin: Springer.
Gendreau, M., Hertz, A., & Laporte, G. (1994). A tabu search heuristic for the vehicle routing problem.
Management Science, 40(10), 1276–1290.
Gendreau, M., Laporte, G., & Séguin, R. (1996). Stochastic vehicle routing. European Journal of Operational
Research, 88(1), 3–12.
Glover, F., & Laguna, M. (1997). Tabu search. Boston: Kluwer Academic.
Guntsch, M., & Middendorf, M. (2001). Pheromone modification strategies for ant algorithms applied to dy-
namic TSP. In E. J. W. Boers et al. (Eds.), Lecture notes in computer science: Vol. 2037. Applications of
evolutionary computing: EvoWorkshops 2001: EvoCOP, EvoFlight, EvoIASP, EvoLearn, and EvoSTIM
(pp. 213–222), Como, Italy, 18–20 April 2001. Berlin: Springer.
150 Swarm Intell (2007) 1: 135–151

Hartl, R. F., Hasle, G., & Janssens, G. K. (2006). Special issue on rich vehicle routing problems: editorial.
Central European Journal of Operations Research, 14(2), 103–104.
Ichoua, S., Gendreau, M., & Potvin, J.-Y. (2003). Vehicle dispatching with time-dependent travel times.
European Journal of Operational Research, 144(2), 379–396.
Kallehauge, B., Larsen, J., & Madsen, O. B. G. (2006). Lagrangian duality applied to the vehicle routing
problem with time windows. Computers and Operations Research, 33(5), 1464–1487.
Kilby, P., Prosser, P., & Shaw, P. (1999). Guided local search for the vehicle routing problem. In S. Voss,
S. Martello, I. H. Osman & C. Roucairol (Eds.), Meta-heuristics: advances and trends in local search
paradigms for optimization (pp. 473–486). Boston: Kluwer Academic.
Kirkpatrick, S., Gelatt, C. D., & Vecchi, M. P. (1983). Optimization by simulated annealing. Science,
220(4598), 671–680.
Kytöjoki, J., Nuortio, T., Bräysy, O., & Gendreau, M. (2007). An efficient variable neighborhood search
heuristic for very large scale vehicle routing problems. Computers and Operations Research, 34(9),
2743–2757.
Labbé, M., Laporte, G., & Mercure, H. (1991). Capacitated vehicle routing on trees. Operations Research,
39(4), 616–622.
Laporte, G., & Louveaux, F. V. (1998). Solving stochastic routing problems with the integer L-shaped method.
In T. G. Crainic & G. Laporte (Eds.), Fleet management and logistics (pp. 159–167). Boston: Kluwer
Academic.
Li, Y., & Chan Hilton, A. B. (2007). Optimal groundwater monitoring design using an ant colony optimization
paradigm. Environmental Modelling and Software, 22(1), 110–116.
Li, F., Golden, B., & Wasil, E. (2005). Very large-scale vehicle routing: new test problems, algorithms, and
results. Computers and Operations Research, 32(5), 1165–1179.
Lourenço, H. R., Martin, O., & Stützle, T. (2003). Iterated local search. In F. Glover & G. Kochenberger
(Eds.), Handbook of metaheuristics (pp. 321–353). Boston: Kluwer Academic.
Maniezzo, V., & Carbonaro, A. (2000). ANTS heuristic for the frequency assignment problem. Future Gen-
eration Computer Systems, 16(8), 927–935.
Mester, D., & Bräysy, O. (2005). Active guided evolution strategies for the large scale vehicle routing problem
with time windows. Computers & Operations Research, 32(6), 1593–1614.
Montemanni, R., Gambardella, L. M., Rizzoli, A. E., & Donati, A. V. (2005). Ant colony system for a dynamic
vehicle routing problem. Journal of Combinatorial Optimization, 10, 327–343.
Osman, I. H. (1993). Metastrategy simulated annealing and tabu search algorithms for the vehicle routing
problem. Annals of Operations Research, 41, 421–451.
Potvin, J.-Y., Xu, Y., & Benyahia, I. (2006). Vehicle routing and scheduling with dynamic travel times. Com-
puters and Operations Research, 33(4), 1129–1137.
Psaraftis, H. (1995). Dynamic vehicle routing: status and prospects. Annals of Operations Research, 61, 143–
164.
Psaraftis, H. (1988). Dynamic vehicle routing problems. In B. L. Golden & A. A. Assad (Eds.), Vehicle
routing: methods and studies (pp. 223–248). Amsterdam: North-Holland.
Reimann, M., Doerner, K., & Hartl, R. F. (2002). A savings based ant system for the vehicle routing problem.
In W. B. Langdon et al. (Eds.), Proceedings of the genetic and evolutionary computation conference
(GECCO-2002) (pp. 1317–1325). San Francisco: Kaufmann.
Reimann, M., Doerner, K., & Hartl, R. F. (2003). Analyzing a unified ant system for the VRP and some
of its variants. In G. Raidl et al. (Eds.), Lecture notes in computer science: Vol. 2611. Applications
of evolutionary computing: EvoWorkshops 2003: EvoBIO, EvoCOP, EvoIASP, EvoMUSART, EvoROB,
and EvoSTIM (pp. 300–310), Essex, UK, 14–16 April 2003. Berlin: Springer.
Reimann, M., Doerner, K., & Hartl, R. F. (2004). D-ants: savings based ants divide and conquer the vehicle
routing problem. Computers and Operations Research, 31(4), 563–591.
Resende, M. G. C., & Ribeiro, C. C. (2003). Greedy randomized adaptive search procedures. In F. Glover &
G. Kochenberger (Eds.), Handbook of metaheuristics (pp. 219–249). Boston: Kluwer Academic.
Savelsbergh, M. W. P. (1985). Local search in routing problems with time windows. Annals of Operations
Research, 4, 285–305.
Stützle, T., & Hoos, H. H. (2000). MAX-MIN ant-system. Future Generation Computer Systems, 16(8),
889–914.
Taillard, È. D., Badeau, E. P., Gendreau, M., Guertin, F., & Potvin, J.-Y. (1997). A tabu search heuristic for
the vehicle routing problem with soft time windows. Transportation Science, 31(2), 170–186.
Toth, P., & Vigo, D. (2001a). Branch-and-bound algorithms for the capacitated VRP. In P. Toth & D. Vigo
(Eds.), The vehicle routing problem (pp. 29–51). Philadelphia: SIAM.
Toth, P., & Vigo, D. (2001b). An overview of vehicle routing problems. In P. Toth & D. Vigo (Eds.), The
vehicle routing problem (pp. 1–26). Philadelphia: SIAM.
Swarm Intell (2007) 1: 135–151 151

Toth, P., & Vigo, D. (2003). The granular tabu search and its application to the vehicle routing problem.
INFORMS Journal on Computing, 15(4), 333–346.
Van Breedam, A. (1996). An analysis of the effect of local improvement operators in genetic algorithms
and simulated annealing for the vehicle routing problem. RUCA Working Paper 96/14, University of
Antwerp, Belgium, 1996.
Zecchin, A. C., Maier, H. R., Simpson, A. R., Leonard, M., & Nixon, J. B. (2007). Ant colony optimization
applied to water distribution system design: comparative study of five algorithms. Journal of Water
Resources Planning and Management, 133(1), 87–92.
Zeimpekis, V., Tarantilis, C. D., Giaglis, G. M., & Minis, I. (2007). Dynamic fleet management—concepts,
systems, algorithms & case studies. Berlin: Springer.