See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/221563433

The Real-Time Vehicle Routing Problem

Conference Paper · January 2007

DOI: 10.1007/978-3-540-77903-2_22 · Source: DBLP

CITATION READS

1 227

2 authors, including:

Knut Richter
Saint Petersburg State University
77 PUBLICATIONS 1,402 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Economy efficiency and the Environment View project

All content following this page was uploaded by Knut Richter on 04 June 2014.

The user has requested enhancement of the downloaded file.

Int. J. Vehicle Information and Communication Systems, Vol. 2, Nos. 1/2, 2009 59

Vehicle routing problem with real-time travel times

Irena Okhrin*
Department of Information & Operations Management,
European University Viadrina,
P.O. Box 1786, 15207 Frankfurt (Oder), Germany
Fax: +49 335 5534 2675 Email: [email protected]
*Corresponding author

Knut Richter
Department of Industrial Management,
European University Viadrina,
P.O. Box 1786, 15207 Frankfurt (Oder), Germany
Email: [email protected]

Abstract: The paper considers a vehicle routing problem with time windows
and real-time travel times. We assume the deployment of an information and
communication system that is based on mobile technologies, which provides
a real-time mobile connection between the dispatching centre and drivers,
allows localising vehicles on road, and gives the online overview over traffic
conditions. We explicitly incorporate the possibility to react to some dynamic
events like traffic impediments and divert a vehicle en route away from its
current destination. We formulated the vehicle routing problem with real-time
travel times as a mixed-integer linear programming model and developed
a genetic algorithm to solve it. Moreover, we performed an extensive
computational study to prove the efficiency of the proposed algorithm on
well-known static benchmarks and to test its performance in dynamic settings.

Keywords: vehicle routing problem; time-dependent; real-time travel times;

mobile information and communication technology; genetic algorithm.

Reference to this paper should be made as follows: Okhrin, I. and Richter, K.

(2009) ‘Vehicle routing problem with real-time travel times’, Int. J. Vehicle
Information and Communication Systems, Vol. 2, Nos. 1/2, pp.59–77.

Biographical notes: Irena Okhrin is an Assistant Professor in Information &

Operations Management at the European University Viadrina in Frankfurt
(Oder). She has received a Diploma in Applied Mathematics and a Master in
Economics degree from Ivan Franko National University of Lviv, Ukraine and
a doctor’s degree in Economics from the European University Viadrina. Her
activities are devoted to the analysis of the influence of wireless technology
on logistics business processes and supply chain management. Her research
interests also include mathematical modelling and new information technology.

Knut Richter is currently a Professor in the Department of Industrial

Management at the European University Viadrina in Frankfurt (Oder). He has
teaching experience as Guest and Associate Professor at universities in
numerous countries like Germany, Ethiopia, Russia, Hungary and Uruguay,
among others. He has been a Coordinator of seven projects on mathematical
modelling and management at the enterprise level. His research and teaching

Copyright © 2009 Inderscience Enterprises Ltd.

60 I. Okhrin and K. Richter

are in the areas of logistics and operations management, mathematical

optimisation, operations research and environmental management. His works
have been published in various academic journals, including the International
Journal of Production Economics, European Journal of Operational Research,
Central European Journal of Operations Research, OR Spectrum and others.

1 Introduction

Vehicle Routing Problems (VRPs) appear in distributing and/or collecting of goods by

commercial or public transport. The aim of a VRP is to determine a route and schedule of
vehicles in order to satisfy customer orders and minimise operational costs. The objective
is usually to minimise travel times, distances, costs or the number of used vehicles while
distributing goods. In addition, many side constrains have to be satisfied, such as vehicle
capacity constrains, route duration constrains, time window constrains, etc.
In the past, vehicles executing routes and dispatchers in the control centre were acting
separately, without or with only little information exchange. The position of vehicles en
route was not known to the dispatcher and it was not always possible to establish a good
connection with drivers. Therefore, all client requests that arrive after vehicles departure
have to be postponed till the next planning period. Recent advances in information
and communication technologies improve dramatically the quality of communication
between drivers and the dispatching office. New customer orders as well as route
changes can now be easily communicated to drivers, thus enhancing service quality and
reducing costs. Moreover, state-of-the-art navigation systems provide real-time traffic
and weather conditions allowing to escape hampered roads.
In the paper, we consider a dynamic vehicle routing problem with time windows
and real-time travel times. To enhance the effectiveness of operations management, we
assume the deployment of an information and communication system that is based on
mobile technologies. It encompasses, first of all, a real-time mobile connection between
the dispatching centre and the drivers. Second, the dispatching office can localise the
vehicles on road via the Global Positioning System (GPS) facilitating fleet utilisation and
increasing productivity. And third, modern technologies are able to provide the overview
over current traffic conditions, what enables the consideration of time-dependent travel
times that are permanently updated during the day. All mentioned possibilities of the
mobile technology allow reacting to some dynamic events like traffic impediments or
new customer request for service and divert a vehicle on road in real time. We formulate
the VRP with online travel time information as a mixed-integer linear programming
model. To solve the problem we develop a genetic algorithm and test its performance on
well-known benchmarks. The achieved results are competitive with best published
solutions.
The contributions of the paper are threefold. First, we discuss the deployment of
modern information and communication systems based on wireless technology in vehicle
routing. Such systems are an indispensable source of real-time data on fleet location
and traffic conditions and serve as a communication medium between vehicles and
dispatchers. They allow using estimations of real travel times in online mode, as provided
by navigation systems, thus eliminating the need to refer to some constant travel times
or base calculations on a step function that is known ex ante. Second, we define a
framework for a VRP with real-time travel times and formulated a mathematical model
 Vehicle routing problem with real-time travel times 61

which allows to consider variable travel times and to capture the dynamics of real-life
systems. By introducing some intermediate artificial point in the road network we
explicitly consider the possibility to divert a vehicle to other destination, if the planned
route is hampered. Thus, vehicles arrive always on time, all customers are served within
provided time windows and costs are minimised. Finally, we developed an algorithm that
can deal with the complexity of the proposed model and test its performance for constant
and variable travel times.
The rest of the paper is organised as follows. Section 2 outlines the application of
mobile technologies to vehicle routing. Section 3 presents the literature review. Further,
the time-dependent VRP with real-time travel times is described in Section 4 and in
Section 5 a corresponding mathematical model is formulated. The genetic algorithm to
solve the developed model is elaborated in Section 6. Finally, Section 7 presents the
computational results for both constant and variable travel times and Section 8 concludes.

2 Mobile technologies in vehicle routing

The novelty of mobile technology is in its ability to provide accurate and critical
information in a timely manner. For systems that are based on dynamic vehicle routing
the flow of information is as important as the flow of physical assets (Cetin and List,
2006). Therefore, their practical implementation is not possible without a real-time, high-
bandwidth information and communication system. Recent developments in wireless
communication technologies make the exchange of information between the dispatching
centre and the vehicles on-road as well as inter-vehicle data exchange especially
cost-effective and uncomplicated. Thus, new instructions for action or recent customer
orders can be sent to drivers at any time, regardless of their location and status. In
addition, the wireless technology helps to capture up-to-the-minute traffic data and sends
it to a powerful workstation. The back-end software analyses the data and transforms it
into the useful information about current traffic conditions and estimations of travel
times. This information is in real-time transmitted to drivers through in-vehicle route
guidance devices and Advanced Traveller Information System (ATIS) facilitating more
informed decision making.
Furthermore, based on signals from satellites, the GPS is used to determine current
vehicle positions, dramatically improving fleet monitoring and management. GPS,
together with Geographic Information Systems (GIS), is a power tool for data
integration, visualisation and analysis. Integrated GPS and GIS are a valuable source of
information for the dispatching centre that can locate the vehicles on road and display
their position on an electronic road map. In addition, dispatchers can determine the
distance between various vehicles of the fleet as well as between a vehicle and its final
destination (Küpper, 2005). This improves fleet utilisation, decreases reaction time, and
gives the overview over routes execution.

3 Literature review

The vehicle routing problem was firstly introduced by Dantzig and Ramser (1959) in
their seminal paper about the delivery of gasoline. Since that time a vast body of research
was devoted to a generic problem and its numerous extensions. Usually the problem
62 I. Okhrin and K. Richter

deals with distribution or collection of goods. However, it is also relevant for delivery of
petroleum and gases, courier services, rescue and repair services, emergency services,
taxi cab and tramp ship operations, combined pick-up and delivery services as well as for
the transportation of elderly and disabled people (Psaraftis, 1995; Ghiani
et al., 2003). A variety of exact and approximate solution methods were developed
for VRP. For the survey of some of them see, for example, Laporte (1992); Golden et al.
(1998); Laporte et al. (2000), etc. Besides this, the two books edited by Golden and Assad
(1988) and Toth and Vigo (2002) were published on the topic.
In practice, vehicle routing problems are mostly dynamic; in literature, however, the
static version of a problem is handled more often. The dynamic or real-time version of a
vehicle routing problem was investigated by Psaraftis (1988, 1995). Earlier works on the
problem also belong to Bertsimas and van Ryzin (1991, 1993). Gendreau and Potvin
(1998) review several variations of a dynamic VRP such as dial-a-ride problems, repair
and courier services as well as express mail delivery. A dynamic pickup and delivery
problem was studied by Swihart and Papastavrou (1999) and recently by Mitrović-Minić
et al. (2004). Attanasio et al. (2004) examined a dynamic dial-a-ride problem. The invited
review on the dynamic vehicle routing was presented by Ghiani et al. (2003). Some
recent works on the topic include, among others, Taniguchi and Shimamoto (2004);
Haghani and Jung (2005); Potvin et al. (2006), etc.
The majority of papers on vehicle routing make an assumption about constant travel
times that are a scalar transformation of distances. This concept, however, does not
represent well ever changing travel times of the real world. The notion of time-dependent
travel times for vehicle routing was introduced by Malandraki (1989) and further
investigated by Malandraki and Daskin (1992) and Malandraki and Dial (1996). The
authors assumed that travel times between two nodes change throughout the day in an a
priori known fashion. This partially captures the predictive variations of the travel time
function. They divide a day into a few time intervals and define travel time variation as a
step function. This function, however, does not satisfy the First-In-First-Out (FIFO)
assumption. That is, if two vehicles depart from the same depot and travel to the same
customer using the same roads, the vehicle that left the depot later may arrive earlier to
its destination compared to its rival.
The situation is improved in the subsequent papers. Ichoua et al. (2003) proposed a
model for the vehicle routing problem based on the time-dependent travel speed which
satisfies the FIFO property. The travel speed changes as time proceeds from one time
interval to another. The work by Chen et al. (2006) also defines travel time between
nodes as a step function, it is however unknown in advanced and may change if an
unexpected event occurs. In addition, Jung and Haghani (2001) and Haghani and Jung
(2005) consider a continuous travel time function and develop a genetic algorithm for the
corresponding VRP. Finally, Donati et al. (2008) present the multi-ant colony system
to simultaneously optimise the number of tours and the total travel time of the
time-dependent VRP in urban context.

4 Problem definition

We consider a dynamic vehicle routing problem with time windows and online travel
time information. The problem is defined on a complete graph G = (V, A), where V = 0,
1, ... , n is the vertex set and A = {(i,j): i,j ∈ V, i ≠ j} the arc set. Vertex 0 represents a
 Vehicle routing problem with real-time travel times 63

depot whilst other vertices represent geographically dispersed customers that have to be
served. A positive deterministic demand is associated with every customer. The demand
of a single customer cannot be split and should be serviced by one vehicle only. Each
customer defines its desired period of time when he wishes to be served. A set of K
identical vehicles with capacity Q is based at the single depot. Each vehicle may perform
at most one route which starts and ends at the depot. The vehicle maximum load may not
exceed the vehicle capacity Q. The objective is to design routes on G such that every
customer belongs to exactly one route and the total travel time of all vehicles is
minimised.
Not all customer requests are known at the beginning of the planning period
(typically, a day). Some of them are revealed in the course of time, when vehicles are
already in transit. This fact makes the problem dynamic or time-dependent. We assume
that the policy of the company is to accept all received orders. Another dynamic
component of the problem is a link travel time function between a pair of nodes. The
frequently considered constant travel time function, proportional to the distance between
the nodes and time independent, is not realistic. It does not represent well the real-world
situation, where travel times vary with time and are subject to numerous factors.
In practice, the link travel time function fluctuates because of changing traffic and
weather conditions like, for example, congestion during rush hours, accidents, etc.
Furthermore, available models for the dynamic vehicle routing usually imply that a
vehicle en route must first reach its current destination and only after that its route can be
changed to incorporate a new customer or to escape a hampered road. It is not possible to
divert a vehicle in transit between two nodes in response to a new customer request or
changing travel times. Exactly on the way to its immediate destination, however, the
vehicle may encounter an unpredicted congestion or other traffic impediment. So, the
vehicle has to wait unreasonably long, endangering the satisfaction of the time window
restrictions of the subsequent clients. Instead of this, it can deviate from its route and
serve other customers in the meantime.
Thanks to mobile technology, we can overcome the mentioned shortcomings and
model vehicle routing in more realistic settings. We assume that there is a real-time
communication between the vehicles and the dispatching centre. Furthermore, the
dispatcher has an overview over current traffic conditions and knows the locations of the
vehicles. Therefore, we explicitly incorporate into the model varying and time-dependent
travel times and the possibility to divert a vehicle en route. To gain an opportunity to
direct a vehicle in transit to another customer, we insert an additional node into the
original road network at the place of vehicle location. Thus we create the so-called
artificial intermediate nodes for these vehicles that at time of routes adjustment are in
transit between two nodes. No demands are associated with these nodes and the vehicles
must immediately leave them. This concept is illustrated in Figure 1. Let us assume that
on its way from node 1 to node 2 vehicle A encounters a hampered road. In order to
avoid unreasonable waiting, we create an artificial intermediate node (denoted by a
square) and run a re-optimisation procedure to find better vehicle routes under given
circumstances. After routes adjustment, vehicle A is diverted to node 3 and node 2 is
visited by vehicle B instead. Hence, we do not commit customers to routes. Clients can
be served by any vehicle that has enough capacity, what increases the flexibility of the
system.
64 I. Okhrin and K. Richter

Figure 1 Anticipated and updated vehicle routes

To better react to changing link travel times or to new customer requests, we use
the concept of time rolling horizon. We formulate a series of mixed-integer linear
programming models, where each model characterises a particular static vehicle routing
problem with heterogeneous fleet at a specific point of time. The re-optimisation
algorithm is then performed on the graph that besides the initial road network includes
also the artificial intermediate nodes. The model is formulated and solved every time τ
when one of the following dynamic events happens:
• A new customer request for service has arrived (if service requests arrive too late to
be served in the current planning period, they are postponed till the next period and
no routes adjustment is undertaken).
• Travel time between a pair of nodes is updated.
• A vehicle is considerably ahead/behind the schedule.
After a new event has arrived, the task of the dispatching centre is to process recent data,
determine new routes, redirect vehicles en routes, update schedules, and finally inform
customers about anticipated arrival times. The routes adjustment procedure should take
into account new travel time information, current vehicle locations and the most recent
service requests that arrived after the latest adjustments in routes. The time between two
subsequent routes adjustments we will call the execution period. Instead of re-optimising
the vehicle routes after the arrival of every dynamic event (what may be impractical), it is
also possible to introduce a tolerance parameter that buffers the incoming events before
further processing. In this case the dispatcher should wait until several dynamic events
occur and only then run the re-routing algorithm. According to Potvin et al. (2006) this
strategy is more beneficial if compared with immediate reaction to every single dynamic
event.

5 The time-dependent vehicle routing problem

5.1 Notations
Data sets
N0 depot
Nd ≡ Ndemand(τ) set of unserved demand nodes at time τ
Ns ≡ Nserviced(τ) set of nodes being served at τ
 Vehicle routing problem with real-time travel times 65

Na ≡ Nartificial(τ) set of artificial intermediate nodes for vehicles

being en route at time τ
N0d ≡ N0 ∪ Nd

Nds ≡ Nd ∪ Ns

N0sa ≡ N0 ∪ Ns ∪ Na

Ndsa ≡ Nd ∪ Ns ∪ Na

N0dsa ≡ N0 ∪ Nd ∪ Ns ∪ Na

K set of vehicles
Constants and parameters
τ time of the last routes determination or update
tij (τ) estimation of link travel time between nodes i and j at time τ
Q vehicle capacity
qi ≤ Q demand of node i
Qk (τ) load of vehicle k at time τ
ei starting time of time window at node i
e0 beginning of the planning period
li ending time of time window at node i
l0 end of the planning period
si ≤ li – ei service time at node i; s0 ≡ 0
ki vehicle that at time τ is located in node i ∈ Ns ∪ Na

M a very big positive number

Decision variables
t
xijk equals one if vehicle k departures from node i to node j at time t,
otherwise equals zero
aik start of service at node i by vehicle k

5.2 Parameters initialisation

At every time moment τ when we construct vehicle routes for the first time or adjust
them in response to some dynamic events, we have to initialise/update input parameters
for the linear programming model. First of all, data sets have to be redefined. Set Nd of
unserved demands contains only these demand nodes that have not yet been serviced by
any vehicle. If at time τ a vehicle is waiting for service or serving a demand node i, then
this node belongs to the set Ns. Please note that Nd ∩ Ns = ∅. If a vehicle at time τ is in
transit between two nodes, we create an artificial intermediate node i ∈ Na for it. In the
next execution period, we consider the nodes from the set Na as the starting nodes for
66 I. Okhrin and K. Richter

the vehicles in transit. If a vehicle has just finished serving a client, we assume that it
is already at the beginning of its journey to the next customer. Therefore, we create
an artificial intermediate node for such vehicle as well. At the beginning of the planning
period, when we create the routes for the first time and τ = e0, we set Ns = Na = ∅.
Next, as the important input parameters for the model, we take the latest estimations
of the link travel times tij(τ) as provided by the deployed navigation system. If they
have changed considerably since the last routes adjustment, it may have a significant
influence on the problem solution. Further, we redefine the parameters Qk (τ), k ∈ K and
ki, i ∈ Ns ∪ Na which represent the load and number of vehicles at time τ and are known
with certainty. Moreover, the time window boundaries for some nodes i ∈ Ns have to be
adjusted too. If ei < τ then we move forwards the starting time of the corresponding time
window and set ei = τ. In addition, we reduce the service time si by the amount of
time that was used at the previous execution period. Besides this, for i ∈ Ns we assume
that the whole amount qi should be unloaded in the next execution period.

5.3 Mathematical model

The time-dependent vehicle routing problem with time windows and online travel
time information is formulated as a mixed-integer linear programming problem. The
mathematical formulation is partly based on the network flow model of Toth and Vigo
(2002, p.158):
min ∑ ∑ ∑ ∑t
i∈ N 0 dsa j∈ N 0 d k ∈K t ≥τ
ij (τ )xijk
t
, (1)

∑ ∑ ∑τ x
j∈ N 0 d k ∈ K t ≥
t
ijk = 1, i ∈ N dsa , (2)

∑ ∑x
i∈ N 0 dsa t ≥τ
t
ijk − ∑ ∑x
i∈ N 0 d t ≥τ
t
jik = 0, j ∈ Nd , k ∈ K , (3)

∑ ∑x
j∈N d t ≥τ
t
0 jk ≤ 1, k ∈ K, (4)

∑ ∑ ∑τ x
i∈ N 0 sa j∈ N 0 d t ≥
t
ijk − ∑ ∑τ x
i∈N dsa t ≥
t
i0k = 0, k ∈ K, (5)

∑
j∈ N 0 d
τ
xijki
= 1, i ∈ Na , (6)

∑
j∈ N 0 d
si
xijki
= 1, i ∈ Ns , (7)

aik + si + tij (τ ) − a jk ≤ (1 − xijk

t
)M , i ∈ N 0 dsa , j ∈ N 0 d , k ∈ K , t ≥ τ , (8)

ei ∑
j∈ N 0 d
t
xijk ≤ aik , i ∈ N ds , k ∈ K , t ≥ τ , (9)

aik ≤ ( li − si ) ∑
j∈ N 0 d
t
xijk , i ∈ N ds , k ∈ K , t ≥ τ , (10)

e0 ≤ a0 k ≤ l0 , k ∈ K, (11)
 Vehicle routing problem with real-time travel times 67

Qk ( τ ) + ∑ q ∑ ∑x
i∈N ds
i
j∈N 0 d t ≥τ
t
ijk ≤ Q, k ∈ K, (12)

t
xijk ∈ {0,1} , i ∈ N 0 dsa , j ∈ N 0 d , k ∈ K , t ≥ τ , (13)

aik ≥ 0, i ∈ N ds , k ∈ K (14)

The objective of this model is to minimise the total travel time over all routes.
Restrictions (2)–(5) constitute the flow conservation constrains. Equation (2) states that
each customer is visited by exactly one vehicle while equation (3) indicates that the same
vehicle enters and leaves a given demand node. According to equations (4) and (5), each
vehicle leaves depot at most once but if vehicle has left depot and is in transit, it must
also return to it. Two next restrictions deal with vehicle starting times. In particularly,
equation (6) specifies that vehicle ki, that at time τ is located at the artificial intermediate
node i, must immediately leave it. According to equation (7) vehicle ki, that is serving or
waiting for service at node i, must leave this node after the end of the service.
In order to guarantee schedule feasibility, the starting times of service at two
connected nodes i and j must satisfy equation (8). Time windows restrictions are
described by equations (9) and (10), where the first constrain prohibits start of service
before the starting time of the corresponding time window and the second constrain
forbids end of service after its closure. Restriction (11) indicates that vehicles may not
leave the depot before the depot opening time and may not return to the depot after the
depot closing time. Restriction (12) states that the cumulative load must not exceed the
vehicle capacity. Finally, equations (13) and (14) are definitional constrains for binary
flow variables and service starting time variables, respectively.

6 Solution method using a genetic algorithm

The time-dependent vehicle routing problem (1)–(14) is a generalisation of the classical

Travelling Salesman Problem (TSP) and thus belongs to the class of NP-complete
problems (Lenstra and Rinnooy Kan, 1981). Exact solution algorithms can solve to
optimality only small instances of the problem, working unreasonably long for larger
problems. Furthermore, the described time-dependent VRP is an asymmetric problem
while a congestion may occur or a road may be blocked in only one direction, whilst the
opposite direction of the same road remains free. In addition, the triangular property is
not satisfied, while the direct link between two nodes may be hampered and a by-pass
is a faster alternative even if one has to travel a longer distance. Hence, exact solution
methods and classical heuristics cannot handle the formulated problem. Therefore, to
solve it we implemented a genetic algorithm metaheuristics. Genetic algorithms were
successfully deployed to VRPs and have proved to produce good quality solutions (e.g.
Thangiah, 1995; Potvin and Bengio, 1996; Taniguchi and Shimamoto, 2004; Ombuki
et al., 2006; Alvarenga et al., 2007; Hanshar and Ombuki-Berman, 2007; Marinakis
et al., 2007, etc.). In particular, Hanshar and Ombuki-Berman (2007) showed that GA
can deliver good results for a dynamic VRP, too.

6.1 Chromosome representation

To apply a genetic algorithm to any specific problem, first a chromosome representation
should be chosen. In our approach, the information about the whole population is
68 I. Okhrin and K. Richter

contained in one three-dimensional matrix. The first dimension of the matrix indexes
various individuals of the population. The rows of the matrix (second dimension)
correspond to different routes, whilst the columns (third dimension) correspond to the
clients visited on each route. The number of routes for each individual, the number of
clients per route, and the fitness values are stored in separate arrays. Such matrix
representation is very convenient, as it allows quick and efficient sorting of the solutions
within the population. This is possible, since the definition of the matrix through pointers
does not require the overwriting of the elements; only redirection of the pointers should
be made instead of this.

6.2 Initial population

To generate an initial population of feasible solutions we sort the customers by the
starting time of the time window. The customer with the earliest starting time is taken as
the first customer in the first route. Further customers are chosen randomly one after the
other and appended to the route until the time schedule and capacity constraints are
satisfied. If after hundred attempts no valid customer for the given route can be chosen,
we initiate a new route. From the rest of the customers we again select the one with
the earliest time window starting time and set this client as the first for the next route.
The procedure is repeated until no unserved customers are left and enough individuals for
the initial population are created.

6.3 Selection criteria

To select a set of parents for further reproduction, we implement a stochastic tournament
selection operator (Dréo et al., 2006, p.88). The core of the operator is a tournament set
which consists of k individuals randomly chosen from the population. In the deterministic
tournament selection the fittest element from the tournament set is selected for the
reproduction. In the stochastic variation of the operator, a random number r ∈ [0, 1] is
generated. If r is less than some number p, called selection pressure, then the fittest
individual from the tournament set is selected for the reproduction. Otherwise, if r ≥ p
then a random individual from the rest of the tournament set is selected. The selection
pressure must be between 0.5 and 1 and shows the bias of the selection operator towards
the fitter solution. It allows some less fit solutions to survive from generation to
generation (Hanshar and Ombuki-Berman, 2007). The individuals that were chosen for
the tournament set are replaced in the population, as this increases the variance of the
process and favours the genetic drift. The pseudocode of the stochastic tournament
selection algorithm is presented in Figure 2.

6.4 Crossover operator

We adopt the special crossover operator developed by Ombuki et al. (2006) that is
particularly suitable for VRP with hard time windows. The authors carried experiments
with two standard crossover operators, namely uniform order crossover and partially
mapped crossover, and found the produced results unsatisfactory. Therefore, they
introduced a new operator, called Best Cost Route Crossover (BCRC). The operator
produces two feasible offspring from two parents p1 and p2 by executing the following
procedure. In the first step, a random route is selected from each parent (route r1 is
 Vehicle routing problem with real-time travel times 69

selected from parent p1 and r2 from p2). Then the customers that belong to the route r2 are
removed from the parent p1. Analogously, the customers belonging to the route r1 are
removed from parent p2. To yield the feasible children, the removed customers should be
selected randomly and re-inserted back into the corresponding solution at the least cost.
For that purpose the algorithm scans all possible locations for insertion and chooses
the feasible ones. The removed customer is then inserted into the place that induces
the minimum additional costs. If no feasible insertion place can be found, a new
route containing the removed customer alone is created and added to the offspring. The
pseudocode of the corresponding BCRC algorithm is listed in Figure 3. The proportion of
the individuals that undergo the crossover operator is determined by a crossover rate; the
rest survives intact.

Figure 2 Tournament selection algorithm

for (i = 1 to Population_Size)
select randomly k individuals from the population //tournament set
generate a random number r between 0 and 1
if r < Selection_Pressure
then select for reproduction the fittest individual
from the tournament set
else select for reproduction a random individual
from the tournament set without the fittest element
end for

Figure 3 Best cost route crossover algorithm

for (i = 1 to Crossove_Rate × Population_Size)
select randomly two parents p1 and p2
// remove the parents from the reproduction set
select randomly one route from each parent, r1 and r2
remove from p1 customers that belong to r2
remove from p2 customers that belong to r1
for every customer in r2 reinsert it into p1 at the best (least cost) place
for every customer in r1 reinsert it into p2 at the best (least cost) place
// if for some customer its reinsertion is infeasible,
// create a separate route for it
end for

6.5 Mutation operator

Finally, a mutation operator is applied to the population to ensure that the algorithm does
not converge prematurely to a local optimum. As mutation introduces a random alteration
to diversify the search, it can be a relatively destructive element, deteriorating the fitness
of the solution. Therefore, the mutation operator is applied to only small fraction of the
70 I. Okhrin and K. Richter

offspring, determined by the mutation rate. We applied a widely-used swap mutation

algorithm, exchanging two customers with similar time windows. The two windows are
considered to be similar if the difference between their starting times is the smallest
(Alvarenga et al., 2007). The swap operator proceeds as follows: A random customer is
selected from a random tour. Then a customer with the similar time window is searched
for and the two customers are exchanged. If the obtained solution is infeasible, the
procedure is repeated. The pseudocode of the swap mutation operator is provided in
Figure 4.

Figure 4 Mutation operator

for (i = 1 to Population_Size)
generate a random number r between 0 and 1
if r < Mutation_Rate then
repeat
select a random customer from a random route
search for a customer from other routes with the similar time
window
swap the two customers
until a feasible solution created
end if
end for

6.6 Construction of a new population

In the new generation, the offspring created by the sequential application of the selection,
crossover and mutation operators completely replace the parents. Only the small numbers
of the worst offspring are left aside and instead of them the best individuals from the
old generation, called elite, are included into the new generation. It ensures that the
best solutions can propagate through generations without the effect of the crossover or
mutation operators. Therefore, the fitness value of the best solution is monotonically
non-decreasing from one generation to another (Dréo et al., 2006, p.91). In the new
generation, however, the elite individuals have to compete with the fittest offspring,
forcing the algorithm to converge towards an optimum.

7 Computational results

7.1 Constant travel time tests

The proposed genetic algorithm was tested in two stages: Stage one with constant travel
times and stage two with variable travel times. Even though the considered problem is
dynamic and time-dependent, the algorithm was initially tested on constant travel time
data to prove its efficiency. For this purpose, we take the Solomon’s benchmark
problems with the long scheduling horizon, namely the sets R2, C2 and RC2 (Solomon,
1987). Test problems with the narrow horizon were discarded as during one trip a vehicle
 Vehicle routing problem with real-time travel times 71

can serve only few customers. So, they are improper for the dynamic problem at hand.
The algorithm was coded in C and run on a 3 GHz Intel Pentium IV machine with 1GB
memory running SuSE Linux 10.2.
The results for the R2 and RC2 problems are presented in Table 1. Column ‘Average
TT’ indicates average travel time calculated over ten runs for each problem instance,
whilst column ‘Best TT’ contains the best found solution. The two columns ‘Best
known’ state the total distance and the number of used vehicles for the best known
published solutions identified by heuristic methods.1 Even though these numbers indicate
the shortest travelled distance and we are minimising the total travel time, this fact does
not influence the results, as at the moment we consider constant travel time problems
and set the vehicle speed equal to one. Hence the travelled time equals the distance. As
benchmark solutions we take the ones produced by heuristics, not by exact algorithms,
while till these days not all Solomon problems are solved to optimality. Furthermore, it is
not always correct to compare exact methods, which produce optimal solutions but need
considerable amount of computational time, and heuristics, which are designed to
produce fast solutions of a good quality.
Table 1 Test results for constant travel times

Genetic algorithm Best known

Problem Average TT Best TT NV TD NV Quality in%
R201 1200.82 1189.28 8 1165.3 8 2.06
R202 1083.56 1073.69 7 1053.9 6 1.88
R203 910.71 897.47 6 910.1 5 −1.39
R204 751.39 739.23 4 770.8 4 −4.10
R205 984.95 975.01 5 972.5 5 0.26
R206 907.44 897.83 5 906.1 3 −0.91
R207 820.64 804.35 4 860.1 4 −6.48
R208 720.98 709.40 4 744.1 4 −4.66
R209 883.59 867.26 5 888.5 5 −2.39
R210 933.94 921.43 6 939.3 3 −1.90
R211 777.45 763.37 4 815.5 5 −6.39
RC201 1309.75 1278.69 8 1274.3 8 0.34
RC202 1130.18 1113.05 7 1119.5 7 −0.58
RC203 976.22 967.82 5 958.0 5 1.02
RC204 807.81 796.58 4 798.4 3 −0.23
RC205 1170.81 1161.55 7 1154.0 7 0.65
RC206 1083.47 1062.75 6 1080.4 6 −1.63
RC207 1000.79 974.50 5 1005.6 7 −3.09
RC208 805.71 790.86 5 820.5 5 −3.61
Note: TT: travel time, NV: number of vehicles, TD: total distance.
72 I. Okhrin and K. Richter

The efficiency of the developed genetic algorithm is presented in column ‘Quality’.

The quality of our solution is defined as the relative deviation from the corresponding
best known solution. The negative value means that our solution is better than the
corresponding best known, while the positive value indicates that our solution is worse.
As can be seen from the table, the received results are comparable with best known so
far. In fact, for eight instances of R2 problems and five instances of RC2 problems we
were able to outperform the best known solutions (emphasised in bold). Figure 5
illustrates the network topology for the received solutions by the example of R211
problem with random geographic data. Concerning the C2 problem set, the produced
results are quite unsatisfactory while the developed algorithm converges prematurely to a
local optimum. Presumably, it is not suitable for clustered problems especially with tight
time windows, or some other sets of genetic algorithm parameters have to be chosen.

Figure 5 Best found solution for R211 problem

7.2 Test results for real-time travel times

The results from the previous section show that the developed genetic algorithm has
proved to be efficient for constant travel times and thus can be further used in the second
stage of our simulation study with real-time travel times. In this stage, we assume that
travel times between a pair of nodes undergo two types of disturbances. On the one hand,
a link travel time function depends on the time of day when a vehicle drives along this
 Vehicle routing problem with real-time travel times 73

link. Thus we capture time dependency due to periodic traffic congestions that is based
on historic data and hence known a priori. On the other hand, we incorporate unpredicted
short-term fluctuations of travel times that occur due to unexpected dynamic events like
accidents.
In practice, actual travel times are provided by an advanced traveller information
system. For our experiment, however, we have to simulate them. Similar to other authors
(e.g. Chen et al., 2006; Donati et al., 2008) we divide the planning period into four
intervals and define travel times as a step function. To introduce the time-dependency of
travel times, we multiply the original values by the coefficients 1.1, 0.9, 1.2 and 0.8,
respectively. The changeovers between intervals are smoothed in order to satisfy the
FIFO condition. The short-term fluctuations of travel times caused by accidents or other
traffic impediments are modelled through random perturbations. They are set equal to
the absolute value of a normal random variable with mean μ = 0 and variation σ = 9
(cf. Potvin et al., 2006).
Further we have to chose how often to update travel times. According to Montemanni
et al. (2005), who carried out extensive tests on both artificial and real data, the best
value for the time slice equals approximately 20 minutes. (The authors consider an
eight-hour working day and divide it into 25 time intervals.) One should not take a very
short time slice, while the routes re-optimisation and driver notification would be
undertaken too often. On the other hand, very large time intervals are also inappropriate.
The vehicles would follow less-than-optimal routes determined for already obsolete
travel times, while the most recent data about traffic conditions would be ignored. We
believe that 10 or 15 minutes is the absolute minimum for the duration of a time slice.
The maximal duration depends on the range and degree of travel time fluctuations and
can vary considerably for different settings.
To simulate the vehicle routing in the dynamic and time-dependent case, we deploy
the concept of events manager (Montemanni et al., 2005; Hanshar and Ombuki-Berman,
2007). The events manager is a separate module that serves as an interface between the
external world and the optimisation procedure and is responsible for routes adjustments
caused by changes in travel times. When a new dynamic event arrives, the events
manager creates a static problem and runs the genetic algorithm to solve it. As inputs the
events manager takes the list of not yet serviced customers, the latest travel time
estimations and the current locations of vehicles. As output, it returns the vehicle routes
that are the best found subject to dynamic travel times. Afterwards, these routes are
transmitted to drivers and anticipated arrival times are communicated to customers.
The pseudocode of the events manager module is presented in Figure 6.
The test results for the real-time case are presented in Table 2. Column ‘Updated
routes’ contains the average travel time value calculated over 20 runs when the routes
re-optimisation was undertaken after every perturbation of the travel time matrix. On the
contrary, column ‘Initial routes’ states the results for the case when travel times are
periodically updated but the routes are not correspondingly adjusted. Consequently, the
vehicles have to follow the initial routes constructed at the beginning of the planning
period. The value difference between the two columns shows that even for small
perturbations of the travel times the periodic routes adjustment leads to better results.
74 I. Okhrin and K. Richter

Figure 6 Events manager

Time = 0
create a list of unserved customer orders
get the latest estimations for the travel times
set vehicle initial locations to depot
repeat
create a static problem
run the genetic algorithm
for (each route from the best found solution)
remove from the unserved list customers that have been served
till the moment Time + Δt
update vehicle capacities
calculate the vehicle locations at the moment Time + Δt
create artificial intermediate nodes for vehicles being on road
at Time + Δt
sum up the travelled till Time + Δt time
end for
update travel times:
→ integrate artificial intermediate nodes
→ multiply travel times by the corresponding coefficient
→ apply random perturbations
update list of rejected customers
Time = Time + Δt
until no unserved customers left

Table 2 Test results for real-time travel times

Updated Initial Rejected Updated Initial Rejected

Problem routes routes in% Problem routes routes in%
R201 1211.06 1218.51 15 RC201 1319.02 1320.84 0
R202 1084.57 1111.92 20 RC202 1148.95 1168.27 5
R203 910.07 929.85 0 RC203 987.44 995.53 20
R204 759.15 757.26 5 RC204 817.23 829.63 10
R205 1003.64 1022.87 10 RC205 1197.12 1179.83 0
R206 910.32 915.43 15 RC206 1098.97 1130.09 15
R207 831.21 836.95 15 RC207 1021.00 1021.01 10
R208 731.18 732.54 20 RC208 825.43 834.00 5
R209 890.04 896.07 10
R210 948.58 956.41 5
R211 794.88 811.11 10
 Vehicle routing problem with real-time travel times 75

Finally, column ‘Rejected’ indicates the fraction of problem instances containing

customers that could not be served in the case without route re-optimisation. (In the case
when route re-optimisation was undertaken after every update of the travel time matrix,
all customers could be served.) The customers must be rejected while the traversed routes
are definitely less-than-optimal while being determined for obsolete travel times. Hence,
the vehicles arrive to the customers after the ending time of the time windows and are not
able to serve them. This result supports the findings of Fleischmann et al. (2004) who
claim that the usage of constant travel times underestimates the route duration of about
10%. Consequently, when solutions for the constant travel time case are deployed in
real-time settings, their optimality and even feasibility are subject to substantial changes.
Therefore, to be able to serve all customers and decrease costs one has to make use of
modern technologies and real-time data and promptly react to ever-changing settings
of the real world.

8 Conclusions

The paper deals with a vehicle routing problem with real-time travel times. In
particularly, we consider a case when there is a mobile connection between the drivers
and the dispatching centre that allows communicating them new instructions to action in
real time. Furthermore, we assume that the central office makes use of the advanced
technologies to gain an overview over variable traffic conditions and locations of
vehicles on road. This fact allows us to investigate time-dependent travel times which are
updated on a permanent basis. Thus we consider the possibility to react to unpredicted
traffic impediments and divert a vehicle en route from its current destination. For the
given problem we formulated a mixed-integer linear programming model and developed
a genetic algorithm to solve it. We performed an extensive computational study in order
to prove the efficiency of the proposed algorithm on the well-known static benchmarks
and to test its performance in dynamic settings. The achieved results are competitive with
the best published solutions.
The computational study that incorporates dynamic or stochastic customers is
postponed for future research. Furthermore, we plan to improve our model by developing
a mechanism that stores elements of best solutions computed for one time slice and takes
this information as an input in the subsequent time slices. Currently for every time slice
the problem is solved from scratch. However, if between time intervals only small
changes of input parameters take place, then the corresponding solutions should not
differ much. Therefore, some sort of long-term memory that contains previous good
solutions or their elements should save computational time and increase efficiency of the
algorithm. What is more, it would be interesting to integrate into the model statistical
data about past customer requests in order to forecast the distribution of future orders in
time and space. And finally, testing the developed framework in real-life settings with
authentic customer orders and travel times would be also of great interest.
76 I. Okhrin and K. Richter

References
Alvarenga, G.B., Mateus, G.R. and de Tomi, G. (2007) ‘A genetic and set partitioning two-phase
approach for the vehicle routing problem with time windows’, Computers & Operations
Research, Vol. 34, pp.1561–1584.
Attanasio, A., Cordeau, J-F., Ghiani, G. and Laporte, G. (2004) ‘Parallel tabu search heuristic for
the dynamic multi-vehicle dial-a-ride problem’, Parallel Computing, Vol. 30, pp.377–387.
Bertsimas, D.J. and van Ryzin, G. (1991) ‘A stochastic and dynamic vehicle routing problem in the
Euclidean plane’, Operations Research, Vol. 39, pp.601–615.
Bertsimas, D.J. and van Ryzin, G. (1993) ‘Stochastic and dynamic vehicle routing in the Euclidean
plane with multiple capacitated vehicles’, Operations Research, Vol. 41, pp.60–76.
Cetin, M. and List, G.F. (2006) ‘Integrated modeling of information and physical flows in
transportation systems’, Transportation Research Part C, Vol. 14, pp.139–156.
Chen, H-K., Hsueh, C-F. and Chang, M-S. (2006) ‘The real-time time-dependent vehicle routing
problem’, Transportation Research Part E, Vol. 42, pp.383–408.
Dantzig, G.B. and Ramser, J.H. (1959) ‘The truck dispatching problem’, Management Science,
Vol. 6, pp.80–91.
Donati, A.V., Montemanni, R., Casagrande, N., Rizzoli, A.E. and Gambardella, L.M. (2008)
‘Time dependent vehicle routing problem with a multi ant colony system’, European Journal
of Operational Research, Vol. 185, pp.1174–1191.
Dréo, J., Pétrovski, A., Siarry, P. and Taillard, E. (2006) Metaheuristics for Hard Optimization,
Springer, Berlin.
Fleischmann, B., Gietz, M. and Gnutzmann, S. (2004) ‘Time-varying travel times in vehicle
routing’, Transportation Science, Vol. 38, pp.160–173.
Gendreau, M. and Potvin, J-Y. (1998) ‘Dynamic vehicle routing and dispatching’, in Crainic, T.G.
and Laporte, G. (Eds): Fleet Management and Logistics, Kluwer, pp.115–126.
Ghiani, G., Guerriero, F., Laporte, G. and Musmanno, R. (2003) ‘Real-time vehicle routing:
solution concepts, algorithms and parallel computing strategies’, European Journal of
Operational Research, Vol. 151, pp.1–11.
Golden, B.L. and Assad, A.A. (1988) Vehicle Routing: Methods and Studies, North-Holland.
Golden, B.L., Wasil, E.A., Kelly, J.P. and Chao, I-M. (1998) ‘Metaheuristics in vehicle routing’,
in Crainic, T.G. and Laporte, G. (Eds): Fleet Management and Logistics, Kluwer, pp.33–56.
Haghani, A. and Jung, S. (2005) ‘A dynamic vehicle routing problem with time-dependent travel
times’, Computers & Operations Research, Vol. 32, pp.2959–2986.
Hanshar, F.T. and Ombuki-Berman, B.M. (2007) ‘Dynamic vehicle routing using genetic
algorithms’, Applied Intelligence, Vol. 27, pp.89–99.
Ichoua, S., Gendreau, M. and Potvin, J-Y. (2003) ‘Vehicle dispatching with time-dependent travel
times’, European Journal of Operational Research, Vol. 144, pp.379–396.
Jung, S. and Haghani, A. (2001) ‘Genetic algorithm for the time-dependent vehicle routing
problem’, Transportation Research Record 1771, pp.164–171.
Küpper, A. (2005) Location-Based Services: Fundamentals and Operation, Wiley, Chichester,
England.
Laporte, G. (1992) ‘The vehicle routing problem: an overview of exact and approximate
algorithms’, European Journal of Operational Research, Vol. 59, pp.345–358.
Laporte, G., Gendreau, M., Potvin, J-Y. and Semet, F. (2000) ‘Classical and modern heuristics for
the vehicle routing problem’, International Transactions in Operational Research, Vol. 7,
pp.285–300.
Lenstra, J.K. and Rinnooy Kan, A.H.G. (1981) ‘Complexity of vehicle routing and scheduling
problems’, Networks, Vol. 11, pp.221–227.
Malandraki, C. (1989) Time Dependent Vehicle Routing Problems: Formulations, Solution
Algorithms and Computational Experiments, PhD Thesis, Northwestern University, Evanson.
 Vehicle routing problem with real-time travel times 77

Malandraki, C. and Daskin, M.S. (1992) ‘Time dependent vehicle routing problems: formulations,
properties and heuristic algorithms’, Transportation Science, Vol. 26, pp.185–200.
Malandraki, C. and Dial, R.B. (1996) ‘A restricted dynamic programming heuristic algorithm for
the time dependent traveling salesman problem’, European Journal of Operational Research,
Vol. 90, pp.45–55.
Marinakis, Y., Migdalas, A. and Pardalos, P.M. (2007) ‘A new bilevel formulation for the
vehicle routing problem and a solution method using a genetic algorithm’, Journal of Global
Optimization, Vol. 38, pp.555–580.
Mitrović-Minić, S., Krishnamurti, R. and Laporte, G. (2004) ‘Double-horizon based heuristic for
the dynamic pickup and delivery problem with time windows’, Transportation Research Part
B, Vol. 38, pp.669–685.
Montemanni, R., Gambardella, L.M., Rizzoli, A.E. and Donati, A.V. (2005) ‘Ant colony system
for a dynamic vehicle routing problem’, Journal of Combinatorial Optimization, Vol. 10,
pp.327–343.
Ombuki, B., Ross, B.J. and Hanshar, F. (2006) ‘Multi-objective genetic algorithms for vehicle
routing problem with time windows’, Applied Intelligence, Vol. 24, pp.17–30.
Potvin, J-Y. and Bengio, S. (1996) ‘The vehicle routing problem with time windows – Part II:
genetic search’, INFORMS Journal of Computing,Vol. 8, pp.165–172.
Potvin, J-Y., Xu, Y. and Benyahia, I. (2006) ‘Vehicle routing and scheduling with dynamic travel
times’, Computers & Operations Research, Vol. 33, pp.1129–1137.
Psaraftis, H.N. (1988) ‘Dynamic vehicle routing problems’, in Golden, B.L. and Assad, A.A. (Eds):
Vehicle Routing: Methods and Studies, North-Holland, pp.223–248.
Psaraftis, H.N. (1995) ‘Dynamic vehicle routing: status and prospects’, Annals of Operations
Research, Vol. 61, pp.143–164.
Solomon, M.M. (1987) ‘Algorithms for the vehicle routing and scheduling problems with time
window constrains’, Operations Research, Vol. 35, pp.254–265.
Swihart, M. and Papastavrou, J.D. (1999) ‘A stochastic and dynamic model for the single-vehicle
pick-up and delivery problem’, European Journal of Operational Research, Vol. 114,
pp.447–464.
Taniguchi, E. and Shimamoto, H. (2004) ‘Intelligent transportation system based dynamic vehicle
routing and scheduling with variable travel times’, Transportation Research Part C, Vol. 12,
pp.235–250.
Thangiah, S.R. (1995) ‘Vehicle routing with time windows using genetic algorithms’, in Chambers,
L. (Ed.): Applications Handbook of Genetic Algorithms: New Frontiers, Vol. 2, CRC Press,
Boca Raton, FL, pp.253–277.
Toth, P. and Vigo, D. (Eds) (2002) The Vehicle Routing Problem, Society for Industrial and
Applied Mathematics, Philadelphia.

Notes
1 The prevailing majority of best-known solutions identified by heuristics are taken from
the article by Alvarenga et al. (2007). The results for the problems R206, R210 and RC204
are taken from the site http://w.cba.neu.edu/~msolomon/problems.htm (accessed on
21 February 07).

View publication stats