European Journal of Operational Research 232 (2014) 464–478

Contents lists available at ScienceDirect

European Journal of Operational Research

journal homepage: www.elsevier.com/locate/ejor

Discrete Optimization

The bi-objective Pollution-Routing Problem

Emrah Demir a, Tolga Bektasß b,⇑, Gilbert Laporte c
a
School of Industrial Engineering, Operations, Planning, Accounting, and Control (OPAC), Eindhoven University of Technology, 5600 MB Eindhoven, The Netherlands
b
School of Management, Centre for Operational Research, Management Science and Information Systems (CORMSIS), University of Southampton, SO17 1BJ Southampton,
United Kingdom
c
Canada Research Chair in Distribution Management, Interuniversity Research Centre on Enterprise Networks, Logistics, and Transportation (CIRRELT), HEC Montréal,
3000 chemin de la Côte-Sainte-Catherine, Montréal H3T 2A7, Canada

a r t i c l e i n f o a b s t r a c t

Article history: The bi-objective Pollution-Routing Problem is an extension of the Pollution-Routing Problem (PRP) which
Received 25 August 2012 consists of routing a number of vehicles to serve a set of customers, and determining their speed on each
Accepted 1 August 2013 route segment. The two objective functions pertaining to minimization of fuel consumption and driving
Available online 9 August 2013
time are conflicting and are thus considered separately. This paper presents an adaptive large neighbor-
hood search algorithm (ALNS), combined with a speed optimization procedure, to solve the bi-objective
Keywords: PRP. Using the ALNS as the search engine, four a posteriori methods, namely the weighting method, the
Vehicle routing
weighting method with normalization, the epsilon-constraint method and a new hybrid method (HM),
Fuel consumption
CO2 emissions
are tested using a scalarization of the two objective functions. The HM combines adaptive weighting with
Multicriteria optimization the epsilon-constraint method. To evaluate the effectiveness of the algorithm, new sets of instances based
Heuristics on real geographic data are generated, and a library of bi-criteria PRP instances is compiled. Results of
extensive computational experiments with the four methods are presented and compared with one
another by means of the hypervolume and epsilon indicators. The results show that HM is highly effective
in finding good-quality non-dominated solutions on PRP instances with 100 nodes.
 2013 Elsevier B.V. All rights reserved.

1. Introduction
Freight transportation lies at the forefront of logistics planning.
Until now, the planning of freight transportation activities has
mainly focused on ways of saving money and increasing profitabil-
ity by considering internal transportation costs only, e.g., fuel cost,
drivers’ wages (see, e.g., Crainic, 2000; Forkenbrock, 1999, 2001).
Freight transportation in the United Kingdom (UK) is responsi-
ble for 22% of the CO2 emissions from the transportation sector,
amounting to 33.7 million tonnes, or 6% of the CO2 emissions in
the country, of which road transport accounts for a proportion of
92% (McKinnon, 2007). The 2008 Climate Change Act commits
the UK to an ambitious and legally binding 80% reduction in green-
house gases (GHG) emissions by 2050, from a 1990 baseline. The
situation in Europe is not much different. According to the TERM
2011 Report published by the European Environment Agency,
transport (including international maritime) contributed 24% of
the overall GHG emissions in the EU-27 countries in 2009, with
road transport accounting for 17% of the total GHG emissions
Vicente (2011). The transportation sector therefore has an impor-
⇑ Corresponding author. Tel.: +44 2380598969.
E-mail addresses: E.Demir@tue.nl (E. Demir), T.Bektas@soton.ac.uk (T. Bektasß),
Gilbert.Laporte@cirrelt.ca (G. Laporte).
tant role to play, as one of the largest GHG contributor, in achieving
reduction targets (Tight, Bristow, Pridmore, & May, 2005).
The carbon dioxide equivalent (CO2e) measures how much glo-
bal warming a given type and amount of GHG may cause, using the
functionally equivalent amount or concentration of CO2 as the ref-
erence. The selection of GHGs to include in the carbon footprint is
an important issue. Wright, Kemp, and Williams (2011) suggest
that a significant proportion of emissions can be captured through
measurement of the two most prominent anthropogenic GHGs,
CO2 and CH4. The emissions of CO2 are directly proportional to
the amount of fuel consumed by a vehicle. This amount is depen-
dent on a variety of vehicle, environment and traffic-related
parameters, such as vehicle speed, load and acceleration (Demir,
Bektasß, & Laporte, 2011). On the other hand, the emissions of
CH4 are a function of many complex aspects of combustion dynam-
ics and of the type of emission control systems used.
Freight companies also generate significant amounts of air pol-
lution besides GHG, including particulate matter (small particles of
dust, soot, and organic matter suspended in the atmosphere), car-
bon monoxide (colorless, odorless, poisonous gas produced when
carbon-containing fuel is not burned completely), ozone (formed
when emissions of nitrogen oxides (NOx) and volatile organic com-
pounds (VOCs) chemically react in the presence of sunlight) and
hazardous air pollutants, also referred to as air toxics (chemicals

0377-2217/$ - see front matter  2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ejor.2013.08.002
 E. Demir et al. / European Journal of Operational Research 232 (2014) 464–478
emitted into the atmosphere that cause or are suspected to cause
cancer or other severe health effects) (PSRC, 2010).
Freight transportation planning has many facets, particularly
when viewed from the multiple levels of decision making. Argu-
ably the most famous problem at this level is the well-known Vehi-
cle Routing Problem (VRP), which consists of determining routes
for a fleet of vehicles to satisfy the demands of a set of customers.
The traditional objective in the standard VRP is to minimize a cost
function which is traditionally considered to be the total distance
traveled by all vehicles. Taking a more explicit look at externalities
of freight transportation, and in particular vehicle routing, Bektasß
and Laporte (2011) introduced the Pollution-Routing Problem
(PRP) which aims at minimizing a total cost function comprising
fuel and driving costs in the presence of time windows.
Most real-world problems involve multiple objectives. In the
context of the PRP, two important objectives should be taken into
account, namely minimization of fuel consumption and the total
driving time. Fuel consumption depends on the energy required
to move a vehicle from one point to another, and is proportional
to the amount of emissions. As discussed in Demir, Bektasß, and
Laporte (2012) for each vehicle there exists an optimal speed yield-
ing a minimum fuel consumption. However, this speed is generally
lower than the speed preferred by vehicle drivers in practice. An-
other important issue in road transportation is time management.
In freight transport terminology, time is money and it is essential
for firms to perform timely deliveries in order to establish and keep
a good reputation. In practice, drivers’ schedules tend to be flexible,
with different numbers of hours worked each day, subject to
driving time regulations. If a saving of one hour can be achieved
on a given vehicle route, this would imply reducing the corre-
sponding driver’s costs by an hour (Fowkes & Whiteing, 2006).
Reduction in time spent on a route can be achieved by traveling
at higher speed, but this, in turn, increases fuel costs and
emissions. Since the two objectives of minimizing fuel and time
are conflicting, the problem requires the use of multi-objective
optimization to allow an evaluation of the possible trade-offs.
In practice, companies would like to minimize their total oper-
ating cost, including those related to fuel and time. However, costs
of fuel, emissions and time might differ from one organization to
another, and in some cases rather significantly. As an example, it
is found that driver costs are paid as hourly wages in some coun-
tries (e.g., UK and USA) whereas they are a monthly salary in oth-
ers. Fuel costs also differ between countries. Finally, carbon costs
vary significantly (£60–£225 per tonne) as discussed in Bektasß
and Laporte (2011). In this paper, we investigate a bi-objective
vehicle routing problem in which one of the objectives is related
to fuel consumption and the other to driving time. The two objec-
tives are treated in their natural units of measurement in order to
eliminate the bias resulting from the cost differences just men-
tioned. The benefit is that managers or users of the approach de-
scribed in the paper can attach cost figures relevant to their
organization and can produce tailored trade-off curves for their
operations.
We propose a solution method based on an enhanced adaptive
large neighborhood search (ALNS) and a specialized speed optimi-
zation algorithm described in Demir et al. (2012). The scientific
contribution of this study is threefold: (i) to introduce of a bi-
objective variant of the Pollution-Routing Problem, (ii) to apply
and test multi-objective techniques to solve the bi-objective PRP,
and (iii) to perform extensive computational experiments using
four a posteriori methods evaluated by means of two performance
indicators. In contrast to existing studies on the ‘‘green’’ VRP (for
which a brief review is presented below), this paper breaks away
from the literature by considering two objectives, one of them
being a comprehensive emissions function incorporating the effect
of load and speed. This study also contributes to the multi-objec-
465
tive optimization literature by presenting a comprehensive com-
parison of four methods on the bi-objective PRP.
The remainder of this paper is organized as follows. In Section 2
we provide a general overview of multi-objective optimization and
we summarize the existing literature on multi-objective and
‘‘green’’ VRPs. Section 3 presents the bi-objective PRP along with
a mathematical programming pformulation. Section 4 includes a
brief description of the heuristic algorithm. Section 5 presents
the generation of the instances and the results of extensive compu-
tational experiments, together with managerial insights. Conclu-
sions are stated in Section 6.
2. Multi-objective optimization
Multi-objective optimization (MOO), also known as multi-
objective programming, multi-criteria or multi-attribute optimiza-
tion, is the process of simultaneously optimizing two or more con-
flicting objectives subject to a number of constraints. In this
section, we consider a MOO problem of the form
ðMOOÞ minimize ff1 ðxÞ; f2 ðxÞ; . . . ; fk ðxÞg ð1Þ
subject to x 2 S ð2Þ
where fk: Rn ! R are k P 2 objective functions to be minimized
simultaneously. The decision variables x = (x1, . . . , xn)T belong to a
non-empty feasible region (set) S # Rn . If there is no conflict be-
tween the objective functions, then a solution in which every objec-
tive attains its optimum values can be found. In this case, no special
methods are needed. To avoid such trivial cases, we assume that no
such solution exists. This means that the objective functions are at
least partly conflicting. They may also be incommensurable, i.e.,
measured in different units (Miettinen, 1999), as is the case in this
paper.
For non-trivial multi-objective problems, one cannot identify a
single solution that simultaneously optimizes every objective.
While searching for solutions, one reaches a point such that, when
attempting to improve an objective, other objectives suffer as a re-
sult. A solution is called non-dominated, Pareto optimal, or Pareto
efficient if it cannot be eliminated from consideration by replacing
it with another solution which improves upon one of the objectives
without worsening another. Finding such non-dominated solu-
tions, and quantifying the trade-offs in satisfying the different
objectives, is the goal of setting up and solving a MOO problem.
The next section presents formal definitions of Pareto optimality.
2.1. Multi-objective optimization methods
In this section, we review several methods for solving MOO
problems and for generating Pareto optimal solutions. General ref-
erences on this topic can be found in Ehrgott and Gandibleux
(2002) and Jozefowiez, Semet, and Talbi (2008a).
Methods for MOO can be classified in various ways. One of them
is based on whether many Pareto optimal solutions are generated
or not, and on the role of the decision maker in solving the MOO
problem (Rangaiah, 2009). One possible classification is where
the methods are initially grouped into two: (i) generating methods
and (ii) preference-based methods. The former group of methods
aims at generating one or more Pareto optimal points without
any prior input from a decision maker. In contrast, the latter uses
extra information from a decision maker as part of the solution
process. Generating methods are further divided into three: (i)
no-preference methods, (ii) a posteriori methods using a scalariza-
tion approach, and (iii) a posteriori method using a multi-objective
approach.
No-preference methods do not require any prior information
and generally yield only one Pareto optimal point. Examples of
466 E. Demir et al. / European Journal of Operational Research 232 (2014) 464–478
such methods include global criterion and the multi-objective bun-
dle method (Wierzbicki, 1999). The second type of a posteriori
methods use a scalarization approach. Scalarization means con-
verting the problem into a single or a family of single objective
optimization problems using a real-valued objective function,
termed the scalarizing function. The weighting and -constraint
methods belong to a posteriori methods based on a scalarization
approach in which a series of scalarized single objective problems
have to be solved to find the Pareto optimal points (Coello, Lamont,
& Van Veldhuizen, 2002; Deb, 2001). The third type of a posteriori
methods use a multi-objective rank trial solutions based on the
objective function values. Examples are the non-dominated sorting
genetic algorithm and ant colony optimization. All a posteriori
methods provide many Pareto optimal solutions to the decision
maker who will then select their preferred one (Fonseca & Fleming,
1995).
Preference-based methods are further divided into two: (i) a
priori and (ii) interactive methods. In the former, preferences of a
decision maker are sought and are then included in the initial for-
mulation of a single objective problem. Some of the a priori meth-
ods are values function methods, lexicographic ordering and goal
programming. The latter requires interaction with the decision
maker during the solution process. Examples are interactive surro-
gate worth trade-off method and the NIMBUS method (Deb &
Chaudhuri, 2007).
This paper focuses on the use of a posteriori methods using sca-
larization of the objective functions. The methods used in this pa-
per are described below.
2.1.1. The weighting method (WM)
In the weighting method, the idea is to associate each objective
function with a weighting coefficient and to minimize the
weighted sum of the objectives. This way, multiple objective
functions are transformed into a single objective function as in
Miettinen (1999). We assume that the weighting coefficients wi
are non-negative for all i = 1, . . . , k. Weights are normalized in such
Pk
a way that i¼1 wi ¼ 1. The MOO is then transformed into the
following problem:
X
k
minimize wi fi ðxÞ
i¼1
subject to x 2 S
Problem (3) and (4) is a single-objective optimization problem
which can be solved by existing methods, such as linear or integer
programming.
2.1.2. The weighting method with normalization (WMN)
This method is an extension of weighting method where the
objective functions are normalized to take values between 0 and
1 (Grodzevich & Romanko, 2006). This is done by using the differ-
ences of optimal function values in the worst (called the Nadir) and
the best (called the Utopia) points, which yield the length of the
intervals in which objective functions vary within the Pareto opti-
mal set. The MOO problem is transformed into the following
problem:
X
k
    ure, these algorithms are able to approximate the whole Pareto
minimize wi fi ðxÞ  zUi = zNi  zUi
i¼1
subject to x 2 S
where zU is the ideal objective vector, i.e., the Utopia point, and zN is
the Nadir point. The ideal point, which optimizes all objective func-
tions, is not normally feasible because of the conflicting objectives
but provides lower bounds for the Pareto optimal set. The ideal
point can be calculated as zUi ¼ fi ðx½i Þ where
zi ¼ arg minx ffi ðxÞ : x 2 Sg. The Nadir point, which corresponds to
the worst objective value for each of the objectives, may be feasible
and provides upper bounds for the Pareto optimal set. It is calcu-
lated as zNi ¼ max16j6k f i ðx½j Þ; 8i ¼ 1; . . . ; k. In practice, these points
can be found by calculating the best and the worst values of each
objective function.
2.1.3. -Constraint method (ECM)
In the -constraint method, only one of the objective functions
is selected to be optimized, while all others are converted into con-
straints by imposing an upper bound. The problem to be solved
then takes the form
minimize f l ðxÞ ð7Þ
subject to f j ðxÞ 6 j 8j ¼ 1; . . . ; k; j – l ð8Þ
x2S ð9Þ
In order to generate as many Pareto optimal solutions as possi-
ble, the right-hand side of constraint (8) is gradually increased by a
small amount and the problem is solved again whenever j is in-
creased. As indicated by Mavrotas (2009), this method offers sev-
eral advantages over WM and WMN. The weighting method may
lead to an extreme solution. In contrast, the -constraint method
is able to produce non-extreme (interior) efficient solutions. More-
over, the computational effort is less than that of the weighting
method. Furthermore, WMN requires the normalization of the
objective functions which may affect the results, whereas this is
not required for the -constraint method. Other extensions of the
-constraint method are discussed by Ehrgott and Ryan (2002)
and Laumanns, Thiele, and Zitzler (2006).
2.1.4. Hybrid method (HM)
This method combines the adaptive weighting and -constraint
methods. The name ‘‘hybrid’’ was suggested by Chankong and
Haimes (1983). The proposed hybrid method takes the form
X
k
 
minimize wi fi ðxÞ  zUi ð10Þ
i¼1
subject to f j ðxÞ 6 j 8j ¼ 1; . . . ; k ð11Þ
ð3Þ x2S ð12Þ
where wi > 0 "(i) i = 1, . . . , k.
ð4Þ
2.2. A brief review of the relevant literature on multi-objective VRPs
In this section, we look at the existing studies on the multi-
objective VRP. The VRP has been the subject of intensive research
efforts both for heuristic and exact optimization approaches. How-
ever, the multi-objective variant of the VRP has not been inten-
sively investigated. A survey of multi-objective on VRPs can be
found in Jozefowiez, Semet, and Talbi (2008b), who present a clas-
sification of objectives related to different aspects of VRPs. These
are: tour (cost, profit, makespan, balance, etc.), nodes/arcs (time
windows, customer satisfaction, etc.), and resources (management
of the fleet, characteristics of the product to collect/deliver, etc.).
Evolutionary algorithms constitute a widely popular approach
in solving multi-objective VRPs. Due to their population-based nat-
ð5Þ
front (or surface) in a single run. An extensive survey on multi-
ð6Þ objective evolutionary algorithms can be found in Zitzler, Deb,
and Thiele (2000), Veldhuizen and Lamont (2000) and Zhou et al.
(2011).
A relevant study is that of Paquette, Cordeau, Laporte, and
Pascoal (2013), who proposed a tabu search algorithm incorporat-
ing a routing cost and quality of service in the dial-a-ride problem,
where the reference point method is used to generate several
 E. Demir et al. / European Journal of Operational Research 232 (2014) 464–478
non-dominated solutions. Another study by Qian, Gribkovskaia,
Laporte, and Halskau (2011) who seek to improve helicopter trans-
portation safety by solving a routing problem with a risk objective
expressed in terms of expected number of fatalities. Finally, Jabali,
Van Woensel, and de Kok (2012) consider a VRP model that takes
into account of travel time, fuel, and CO2 emissions costs in a
time-dependent context, where the latter are estimated using
emission functions provided in the MEET report (Hickman, Hassel,
Joumard, Samaras, & Sorenson, 1999). The authors describe a tabu
search algorithm for the problem and show, through computa-
tional experiments, that limiting vehicle speeds to a certain extent
is effective in reducing emissions although costly in terms of total
travel time.
2.3. A brief review of the relevant literature on ‘‘green’’ VRPs
This section first reviews the methodological on the green VRP,
followed by case studies.
2.3.1. Methodological studies
A relevant study is by Maden, Eglese, and Black (2009), who
look at the vehicle routing and scheduling problem to minimize
the total travel time under congestion. The authors take into ac-
count regular congestion due to volume of traffic, and long-term
road works, which can be predicted from historical data. They pro-
pose a heuristic algorithm to consider time-varying travel times. In
order to reflect the characteristics of a real-life problem, their algo-
rithm allows the time required to travel between locations to vary
according to the time at which the journey starts (Eglese, Maden, &
Slater, 2006). The authors consider the current driving legislation
such that: (i) there must be a driving break of 45 minutes every
4.5 hours, (ii) if the total working time is greater than 6 hours then
a 30 minutes break must be taken and (iii) if the total working time
is greater than nine hours then a 45 minutes break must be taken.
Their results suggest that the proposed approach may yield up to
7% savings in CO2 emissions.
In the context of CO2 minimization, Urquhart, Scott, and Hart
(2010b) study the Traveling Salesman Problem (TSP) to identify
tours with low CO2 emissions. The authors examine two different
fuel emission models, which are a power based instantaneous fuel
consumption model introduced by Bowyer, Biggs, and Akçelik
(1985) and a simpler spreadsheet based model from the UK Na-
tional Atmospheric Emission Inventory. The authors use the latter
model in their solution approach. Computational results on six ran-
domly generated instances, each with between 10 and 30 delivery
points, suggest that only a small improvement can be achieved
using the fuel emission model because of the inadequacy of the
simple spreadsheet emission model. Another work by Urquhart,
Hart, and Scott (2010a) studies the VRPTW to build low CO2 solu-
tions using evolutionary algorithms, using an instantaneous fuel
consumption model (Bowyer et al., 1985). The authors look at
the trade-off between CO2 savings, distance and the required num-
ber of vehicles. Their results indicate that savings of up to 10% can
be achieved, depending on the problem instance and the ranking
criterion used in the evolutionary algorithm. In a related study
by Scott, Urquhart, and Hart (2010), the authors study the effect
of topology and payload on CO2 optimized vehicle routing, by using
the COPERT estimation model which is based on the average speed
of a vehicle and includes payload and correction factors for heavy
goods vehicles (Ntziachristos & Samaras, 2000). They work on two
different sets of problem data: delivery of groceries to households
and delivery of paper from a central warehouse. They also work on
several TSP instances randomly chosen from problem datasets, and
assign average speeds for each road category. Their results suggest
that the effect of gradient and payload are highly dependent on the
mixture of the problems studied. The difference in CO2 emissions
467
between the solutions is found to be less than 2.1% for the COPERT
model as in stated in the literature.
An emissions VRP (EVRP), which is an extension of the TDVRP,
was introduced by Figliozzi (2010), who also describes a formula-
tion and solution approaches for the problem. The objective of the
EVRP is the minimization of emission costs, which are proportional
to the amount of GHG emitted which, in turn, is a function of speed
and distance traveled. In the solution approach, a partial EVRP is
first solved to minimize the number of vehicles by using a TDVRP
algorithm, and emissions are then optimized subject to a fleet size
constraint. The departure times are also optimized using the pro-
posed algorithm for any pair of customers. Computational results
obtained on Solomon’s test instances suggest that route character-
istics and the type of the dominant constraint both play a signifi-
cant role in the results. In a related study, Figliozzi (2011)
focused on the analysis of CO2 emissions for different levels of con-
gestion and time-definitive customer demands. Travel time data
from an extensive archive of freeway sensors were used to analyze
the impacts of congestion on commercial vehicle emissions. Com-
putational results suggest that congestion impacts on commercial
vehicle emissions are highly significant, and public agencies and
highway operators must carefully study the implications of poli-
cies that limit travel speeds or increase speed limits since they
may lead to an increase in CO2 emissions. According to Figliozzi
(2011), if travel speed is reduced to an optimal level from an emis-
sions perspective, emissions can be reduced without a significant
increase in fleet size or distance traveled.
Bektasß and Laporte (2011) introduced the Pollution-Routing
Problem (PRP), which considers minimization of fuel cost and dri-
ver costs. In estimating pollution, the authors investigate factors
such as speed, load, and time windows, using the emissions func-
tions proposed by Barth, Younglove, and Scora (2005). In these
functions, the engine-out emission rate is directly related to the
rate of fuel use. Bektasß and Laporte (2011) assume that in a vehicle
trip all parameters will remain constant on a given arc, but load
and speed may change from one arc to another. Their model
approximates the total amount of energy consumed on the arc,
which directly translates into fuel consumption and into GHG
emissions. The authors have run experiments on three classes of
10 instances with 10, 15 and 20 nodes, where the nodes represent
UK cities. They have analyzed cases where customer demands are
generated randomly based on a discrete uniform distribution, as
well as the effects of variance in demand, of vehicle type and of
time windows. Computational results reported by the authors sug-
gest that, by using the proposed approach, energy savings can be
up to 10% when time windows are in place, and up to 4% when
the demand variation is high.
2.3.2. Case studies
Tavares, Zsigraiova, Semiao, and da Graça Carvalho (2008)
looked at optimization of routing networks for waste transporta-
tion. The authors proposed the use of geographic information sys-
tems (GIS) 3D route modeling to optimize the route with an aim to
minimize fuel consumption in different municipalities of the island
of Santo Antao of Cape Verde. Their model takes into account of
both the road angle and the vehicle load. Their findings indicate
that optimization of fuel consumption yields up to 52% savings in
fuel when compared to routes with the shortest distance, even if
this implies increasing the travel distance by 34%. Another work
by Tavares, Zsigraiova, Semiao, and Carvalho (2009) is on the opti-
mization of municipal solid waste collection routes to minimize
fuel consumption using 3D GIS modeling. The authors make use
of the COPERT model described by Ntziachristos and Samaras
(2000). For the case of the city of Praia, their approach reduces
the traveled distance by 29% and fuel consumption by 16%. For
the case of the Santiago island, savings in fuel consumption is
468 E. Demir et al. / European Journal of Operational Research 232 (2014) 464–478

found to be 12%. Another real-life application was presented by
Apaydin and Gonullu (2008). The authors attempt to control emis-
sions in the context of route optimization of solid waste in Trabzon,
Turkey, with a constant emission factor to estimate fuel consump-
tion. The authors aim to minimize the distance traveled by the
trucks. Their results suggest that the route distance and time can
be decreased by 24.6% using their approach, with implications of
reducing CO2 emission by 831.4 grams on each route.
A heterogeneous fleet to reduce emissions was studied by
Pitera, Sandoval, and Goodchild (2011). The authors gave a mathe-
matical formulation and proposed metaheuristics algorithm.
According to their results, emissions and costs can be reduced by
an average of almost 6% and 9%, respectively. They have also
looked at the service quality. For example, reducing delivery fre-
quency to once a day leaded to emissions and cost savings close
to 35% and 3%, respectively. The last case study reviewed here is
by Ubeda, Arcelus, and Faulin (2011), who investigated the
environmental effects of routing in Eroski, Spain. The authors com-
pared four different approaches, namely the current approach,
rescheduling (CVRP), backhauling (VRPB) and green VRP. They
used a matrix of emissions based on the estimation of CO2 emitted
between each link as described by Palmer (2007). Their results sug-
gest that the implementation of the green routing approach has
benefits from economic and ecological perspectives. In terms of
the green aspect of routing, the results indicate that savings of
13.06% in distance and of 13.15% in emissions can be achieved
using the green VRP approach.
3. The bi-objective Pollution-Routing Problem
We now formally describe the bi-objective PRP. This problem is
defined on a complete directed graph G ¼ ðN ; AÞ where
N ¼ f0; . . . ; ng is the set of nodes, 0 is a depot and
A ¼ fði; jÞ : i; j 2 N and i – j} is the set of arcs. The distance from
node i to node j is denoted by dij. A fixed-size fleet of m vehicles,
each of capacity Q, is available to serve the nodes. The set
N 0 ¼ N n f0g is a customer set, and each customer i 2 N 0 has a
non-negative demand qi as well as a time interval [ai, bi] in which
service must start; early arrivals to customer nodes are permitted
but a vehicle, arriving early must wait until time ai before service
can start. The service time of customer i is denoted by ti. The
PRP, proposed by Bektasß and Laporte (2011) and further studied
by Demir et al. (2012), is a single-objective optimization problem.
The bi-objective PRP consists of constructing routes for a set of
vehicles to meet the demands of all customers in a way that all
vehicles depart from and return to the depot node, no vehicle car-
ries load more than its capacity and each customer is visited within
its respective time window. There are two conflicting objectives in
the PRP, namely fuel consumption and total driving time. These
objectives are described in greater detail below, following the def-
inition of the decision variables.
3.2. The fuel consumption objective
The fuel consumption objective is based on the comprehensive
emissions model described by Barth et al. (2005), Scora and Barth
(2006), and Barth and Boriboonsomsin (2008), which is an instan-
taneous model estimating fuel consumption for a given time in-
stant. According to this model, the fuel rate is given by
FR ¼ nðkNV þ P=gÞ=j ð13Þ
where n is the fuel-to-air mass ratio, k is the engine friction factor, N
is the engine speed, V is the engine displacement, and g and j are
constants. The parameter P is the second-by-second engine power
output (in kilowatt), and can be calculated as
P ¼ Ptract =gtf þ P acc ð14Þ
where gtf is the vehicle drive train efficiency, and Pacc is the engine
power demand associated with running losses of the engine and the
operation of vehicle accessories such as air conditioning. The
parameter Ptract is the total tractive power requirements (in kilo-
watt) placed on the wheels:
Ptract ¼ ðM s þ Mg sin h þ 0:5C d qAv 2 þ MgC r cos hÞv =1000 ð15Þ
where M is the total vehicle weight (kilogram). More specifically,
M = w + f, where w is the curb weight (i.e., the weight of an empty
vehicle) and f is the vehicle load. Furthermore, v is the vehicle speed
(meter/second), s is the acceleration (meter/square second), h is the
road angle, g is the gravitational constant, and Cd and Cr are the
coefficients of the aerodynamic drag and rolling resistance, respec-
tively. Finally, q is the air density and A is the frontal surface area of
the vehicle. For a given arc (i, j) of length d, let v be the speed of a
vehicle speed traversing this arc. To simplify notation, let k = n/jw
and c = 1/1000ntfg be constants and w is the conversion factor of
fuel from gram/second to liter/second, and a = s + gsin h + gCrcos h
be a vehicle-arc specific constant and b = 0.5CdqA be a vehicle-spe-
cific constant.
If one omits the indices (i, j) on the variables v, d, f and a to sim-
plify the presentation, the fuel consumption of a vehicle on this arc
as a function of speed v and load f can be expressed as follows:
Fðv ; MÞ ¼ kðkNV þ wcav þ caf v þ bcv 3 Þd=v ð16Þ
Fig. 1 presents how fuel consumption (liter/100 kilometer) changes
with respect to speed (kilometer/hour) for an empty and fully
loaded medium-duty vehicle using the expression (16).
Fig. 1 shows that the vehicle speed that minimizes fuel con-
sumption is around 55 kilometer/hour using the assumed values,
which holds for both when the vehicle is empty and loaded.

3.1. Decision variables

In order to model the PRP, we use a discretized speed function

defined by R non-decreasing speed levels v  r ðr ¼ 1; . . . ; RÞ. The deci-
sion variables are defined as follows: zrij is a binary variable equals
1 if arc ði; jÞ 2 A is traversed at a speed level r, and 0 otherwise; xij is
a binary variable equal to 1 if and only if arc (i, j) appears in the
solution; fij is a non-negative continuous variable representing
the total amount of flow on each arc ði; jÞ 2 A; yj is a non-negative
continuous variable representing the time at which service starts
at node j 2 N 0 . Fig. 1. Fuel consumption vs speed.
 E. Demir et al. / European Journal of Operational Research 232 (2014) 464–478 469

X
The mathematical representation of the fuel consumption x0j ¼ m ð23Þ
objective is shown below: j2N
X
xij ¼ 1 8i 2 N 0 ð24Þ
X X
R
minimize kNVkdij zrij =v r ð17Þ j2N
X
ði;jÞ2A r¼1 xij ¼ 1 8j 2 N 0 ð25Þ
X i2N
þ wckaij dij xij ð18Þ X X
ði;jÞ2A
fji  fij ¼ qi 8i 2 N 0 ð26Þ
j2N j2N
X
þ ckaij dij fij ð19Þ qj xij 6 fij 6 ðQ  qi Þxij 8ði; jÞ 2 A ð27Þ
ði;jÞ2A X
yi  yj þ t i þ dij zrij =v r 6 K ij ð1  xij Þ 8i 2 N ; j 2 N 0 ; i – j ð28Þ
X X
R
2
r2R
þ bckdij v
zrij ð  r Þ ð20Þ ai 6 yi 6 bi 8i 2 N 0 ð29Þ
ði;jÞ2A r¼1 X
yj þ tj  sj þ dj0 zrj0 =v r 6 Lð1  xj0 Þ 8j 2 N 0 ð30Þ
The objective function (17)–(20) is derived from (16). For further r2R
details, the reader is referred to Demir et al. (2012). X
R
zrij ¼ xij 8ði; jÞ 2 A ð31Þ
3.3. The driving time objective r¼1

xij 2 f0; 1g 8ði; jÞ 2 A ð32Þ

The driving time objective is the sum of the total journey time
of all routes starting and ending at the depot. This time is equal to
the arrival time to the depot under assumption that vehicles start
their journey at time zero. The variable sj represents the total time
spent on a route that has a node j 2 N 0 as last visited before
returning to the depot. The mathematical representation of the
driving time objective is
X
minimize sj
j2N 0
This objective measures the total driving time. The total time
spent on a route where customer j 2 N 0 is visited last before
returning to the depot can be calculated as
sj ¼ cj þ tj þ dj0 =v r
where cj is the waiting time at node j, tj is the service time at node j,
and (dj0/vr) is the journey time from last node of a route to the
depot.
fij P 0 8ði; jÞ 2 A ð33Þ
yi P 0 8i 2 N 0 ð34Þ
zrij 2 f0; 1g 8ði; jÞ 2 A; r ¼ 1; . . . ; R ð35Þ
Constraints (23) state that each vehicle must leave the depot. Con-
straints (24) and (25) are the degree constraints which ensure that
each customer is visited exactly once. Constraints (26) and (27) de-
ð21Þ fine the arc flows. Constraints (28)–(30), where Kij = max{0, bi + ti + -
dij/lij  aj}, and L is a large number, enforce the time window
restrictions. Constraints (31) ensure that only one speed level is se-
lected for each arc and zrij ¼ 1 if xij = 1.
The PRP is NP-hard since it is an extension of the classical VRP.
Bektasß and Laporte (2011) have shown that even a simplified ver-
sion of this problem cannot be solved optimally for mid-size in-
ð22Þ
stances. For this reason, heuristics are needed to obtain good-
quality solutions within short computational times, one which
we use in this study.

4. A bi-objective adaptive large neighborhood search algorithm

3.4. Constraints with speed optimization procedure

The constraints of the integer programming formulation of the To solve the bi-objective PRP, we use an enhanced version of the
bi-objective PRP are similar to those given in Bektasß and Laporte ALNS algorithm introduced in Demir et al. (2012). The ALNS algo-
(2011) and Demir et al. (2012), and are shown below: rithm is used as a search engine to find a set of non-dominated

Table 1
Parameters used in the computational tests.

Notation Description Typical values

w Curb-weight (kilogram) 6350
n Fuel-to-air mass ratio 1
k Engine friction factor (kilojoule/revolution/liter) 0.2
N Engine speed (revolution/second) 33
V Engine displacement (liter) 5
g Gravitational constant (meter/square second) 9.81
Cd Coefficient of aerodynamic drag 0.7
p Air density (kilogram/square meter) 1.2041
A Frontal surface area (square meter) 3.912
Cr Coefficient of rolling resistance 0.01
ntf Vehicle drive train efficiency 0.4
g Efficiency parameter for diesel engines 0.9
fc Fuel and CO2e emissions cost per liter (£) 1.4
fd Driver wage per (£/second) 0.0022
j Heating value of a typical diesel fuel (kilojoule/gram) 44
w Conversion factor (gram/second to liter/second) 737
lij Lower speed limit on arc (i, j) (meter/second) 5.5 (or 20 kilometer/hour)
uij Upper speed limit on arc (i, j) (meter/second) 27.8 (or 100 kilometer/hour)
470 E. Demir et al. / European Journal of Operational Research 232 (2014) 464–478

solutions. Here, we only briefly describe the ALNS for reasons of Table 2
The general structure of 100-node instances.
space but refer the interested reader to Demir et al. (2012) for fur-
ther details. Instance Average # Time window Service Load
The ALNS metaheuristic is an extension of the large neighbor- sets of time (kilogram)
vehicles (second)
hood search (LNS) heuristic first proposed by Shaw (1998), and
Lower Upperbound
based on the idea of modifying an initial solution by means of de- bound (second)
stroy and repair operators. If the new solution is better than the (second)
current best solution, it replaces it and use as an input to the next 1 5 0 32,400 300 180
iteration. The ALNS heuristic was put forward by Ropke and Pising- 2 5 600–2400 27,000– 300 180
er (2006) to solve variants of the VRP. Rather than using one large 32,400
neighborhood as in LNS, it applies several removal and insertion 3 5 0 32,400 300 130–230
4 5 600–2400 27,000– 300 130–230
operators to a given solution. Insertion operators are used to repair 32,400
a partially destroyed solution by inserting the nodes in the removal 5 10 0 32,400 600 360
list back into the solution. These operators insert the removed 6 10 600–2400 27,000– 600 360
nodes back into the existing routes when feasibility with respect 32,400
7 10 0 32,400 600 310–410
to the capacity and time windows can be maintained, or they cre-
8 10 600–2400 27,000– 600 310–410
ate new routes. The neighborhood of a solution is obtained by 32,400
removing some customers from the solution and reinserting them 9 20 0 32,400 900 720
back. Our implementation uses the classical Clarke and Wright 10 20 600–2400 27,000– 900 720
(1964) heuristic to construct an initial solution. Furthermore, we 32,400
11 20 0 32,400 900 670–770
use 12 removal and five insertion operators in the ALNS algorithm, 12 20 600–2400 27,000– 900 670–770
which are selected dynamically in the algorithm according to their 32,400
past performance. To this end, each operator is assigned a score 13 5–20 600–2400 27,000– 300–900 180–760
which is increased whenever it improves the current solution. 32,400
The new solution is accepted if it satisfies a criterion defined by
the simulated annealing (Kirkpatrick, Gelatt, & Vecchi, 1983) used
as local search framework applied at the outer level.
Although genetic algorithms appear to be the standard ap- service times and load are randomly generated within these inter-
proach of multi-objective optimization for vehicle routing, our vals. Each instance set is of a different nature, characterized by the
choice behind the use of the ALNS is primarily due to its effective- average number of vehicles (minimum number required based on
ness as was shown by Ropke and Pisinger (2006), Demir et al. load), time windows (loose or tight) and load (homogeneous or
(2012). At each iteration of the algorithm a speed optimization pro- heterogeneous). All instances are available for downloading from
cedure (SOP) is applied. Given a vehicle route, the SOP consists of http://www.apollo.management.soton.ac.uk/prplib.htm.
computing the optimal speed on each arc of the route in order to The proposed algorithm was implemented in C. All experiments
minimize an objective function comprising fuel consumption and were conducted on a server with 3 gigahertz speed and 1 gigabyte
driving time. This is the same algorithm as described in Demir RAM. A preliminary analysis was conducted to fine-tune the
et al. (2012) to which interested readers may refer. parameters. No claim is made that our choice of parameter values
is the best possible. However, the settings used worked well in our
preliminary analysis. The detailed values of all ALNS parameters
5. Computational results are presented in Demir et al. (2012), with the exception of the
new settings given in Table 3.
This section presents the results of extensive computational We ran the algorithm, N1 times each with N2 iterations for each
experiments performed to assess the performance of the multi- instance. The scores of the operators were updated every N3 itera-
objective methods using our ALNS algorithm with speed optimiza- tions. The removable nodes are randomly chosen between s = 4
tion procedure. We first describe the generation of the test in- and s ¼ 16 at each iteration.
stances, the parameters and the quality indicators used to assess
the performance of the proposed methods. We then present the
computational results. The parameters used in the experiments
5.2. Bi-objective solution methods
are given in Table 1.
For the parameters listed in Table 1, lower and upper speed lim-
We have tested the following four methods. In each, the ALNS is
its are initially set to 20 kilometer/hour and 100 kilometer/hour,
used as the search engine to find and store the non-dominated
respectively. At first glance, these limits may not be seen as being
solutions. (i) WM: Here the objective is to minimize the sum of a
very practical. To shed light on this question, we conduct in Sec-
weighted bi-objective function. The weights are increased from
tion 5.5.1 some experiments with these settings. We show that
zero to one in increments of 0.1. The aggregated objective function
the lower speed limit does not significantly influence the resulting
is calculated as w f1 + (1  w) f2 where, f1 is the fuel consumption
speeds.
(in liter) and f2 is the driving time (in hour). (ii) WMN: The objec-
   
tive functions are normalized using Nadir zNi and Utopian zUi
5.1. Generation of the test instances points (i = 1, 2). The aggregated objective function is calculated
       
as w f1  zU1 = zN1  zU1 þ ð1  wÞ f2  zU2 = zN2  zU2 , where f1 and
For the experiments, 13 sets of 10 instances each were gener- f2 are defined as above. (iii) ECM: The algorithm is first run to find
ated, resulting in a total of 130 instances. Each instance has 100 the value 2 which is the minimum value f2 attains at the end of N2
nodes, which represent randomly selected cities from the UK, iterations. The arrival time to the depot is then fixed to 2. The
and uses real road distances. Table 2 provides the design of each algorithm is then run to minimize f1, and 2 is increased by 1 at
instance set. This table presents the number of vehicles, the lower every iteration. (iv) HM: The algorithm starts as in ECM. However,
and upper bounds of time windows, the service times and the load the objective function
  is calculated
 as the weighted sum of two
intervals for each instance set. Data pertaining to time windows, functions as w f1  zU1 þ ð1  wÞ f2  zU2 . The weight w is updated
 E. Demir et al. / European Journal of Operational Research 232 (2014) 464–478 471

during the algorithm. To update the weights during the search, we 5.3.2. Epsilon indicator
have used the same procedure as in Paquette et al. (2013): The epsilon indicator I ðS; RÞ was proposed by Zitzler et al.
  (2003). An illustration of the epsilon indicator in two dimensions
wi ¼ wi1 fi ðxÞ  zUi ði ¼ 1; . . . ; kÞ ð36Þ is given in Fig. 2(b). The epsilon indicator relies on the concept of
, epsilon-dominance which is explained below:
X k
wh ¼ wh wi ðh ¼ 1; . . . ; kÞ ð37Þ
i¼1 Definition 1. Additive -dominance relation. Let xi ; xj 2 S; xi is said
to additively -dominate xj, and written as xj  +xi (or equivalently
where wh is the normalized weight of the objective function h. f(xi) 6  + f(xj)).

5.3. Solution quality indicators Definition 2. Multiplicative -dominance relation. Let xi ; xj 2 S; xi

The performance assessment of techniques in multi-obective
optimization is less straightforward than in single objective opti-
mization. Whereas the output of a single objective function can
be compared directly with lower or upper bounds, the output of
multi-objective optimization is a set of solutions approximating
the Pareto optimal front. Zitzler, Thiele, Laumanns, Fonseca, and
da Fonseca (2003) present a review of the existing quality assess-
ment indicators. We use two of these to compare our four meth-
ods, namely the hypervolume indicator and the -indicator
(Fig. 2). These are described in more detail below. We also use
the number of Pareto solutions found as an additional way of
assessing performance.
is said to multiplicatively -dominate xj, and written as xj  xi
(or equivalently f(xi) 6 f(xj)).
Of these two definitions, the multiplicative one is the most
commonly used one in the literature. The -indicator is based on
the weakly dominated dominance relation. For any method, the
smaller is the  value, the better is the performance of the method.
In two dimensions, the -indicator can be calculated as
I ðS; RÞ ¼ maxr2R mins2S ðs; rÞ ¼ maxffi ðsÞ  ri j1 6 i 6 kg, where ri
is the ith component of the objective vector r, and R is the refer-
ence set defined here as the union of all known Pareto optimal
solutions.
5.4. Results of the methods on PRP instances

We now present the computational results obtained on PRP in-

5.3.1. Hypervolume indicator
stances. Table 4 provides, for each method tested, the average CPU
This metric was proposed by Zitzler and Thiele (1998a, 1998b).
time (in seconds) required to solve all instances of each set. All val-
The hypervolume indicator Ihv ðHÞ computes the volume of a given
ues are averaged over the 10 instances of each set. In this table, the
region H. It is depicted in Fig. 2(a) for the bi-objective case. In this
instance sets are grouped into five. Group I is characterized by
case, each point in the approximation set forms a rectangle shown
homogeneous loads and loose time windows, whereas Group II
by the shaded area with respect to a reference point (generally the
has homogeneous loads but tight time windows. Group III in-
Nadir point) that lies beyond the bounds of the approximation set.
stances have heterogeneous loads and loose time windows. Finally,
The hypervolume indicator is the area of the union of all rectangles
Group IV and V instances have heterogeneous loads and tight time
(see Fig. 2(a)). The larger the value of the indicator (area), the bet-
windows.
ter is the set of solutions.
As can be seen from Table 4, the methods are fast. Instances of
100 nodes are solved within 155 seconds on average for the HM
Table 3 method, based on around 110,000 iterations in total. The data in
Parameters used in the ALNS heuristic. Table 4 reveal that for any type of instance, the overall CPU times
Description Typical values required by HM is smaller than that of the other methods. The ECM
method has the second best performance in terms of CPU times.
Number of loops (N1) 11
Total number of iterations (N2) 10,000
The average computation times of two weighting methods (WM
Number of iterations for roulette wheel (N3) 200 and WMN) are almost the same. This is because these two methods
Lower limit of removable nodes (s) 4 require a similar computational effort. The average CPU time of
Upper limit of removable nodes ðsÞ 16 Group I is around 138 seconds for HM whereas Group II is around
The increase rate of the ECM (1) 300 seconds 148 seconds. The difference is due to the heterogeneous load dis-
tribution of customers. The effect of time windows can be seen

Fig. 2. Solution quality indicators.

472 E. Demir et al. / European Journal of Operational Research 232 (2014) 464–478

Table 4
Average CPU times of the four solution methods (in seconds).

Instance groups Instance sets WM WMN ECM HM

I 1 198.4 199.5 196.6 198.7
5 133.8 132.9 127.8 123.9
9 120.8 119.4 107.6 93.9
Average 151.0 150.6 144.0 138.8
II 2 212.2 218.8 210.6 210.0
6 148.2 147.2 135.5 131.6
10 127.0 126.9 114.5 102.3
Average 162.5 164.3 153.5 148.0
III 3 237.0 240.9 235.8 232.7
7 156.4 156.6 147.7 144.5
11 132.3 132.0 127.1 119.7
Average 175.2 176.5 170.2 165.6
IV 4 246.4 248.8 248.1 245.3
8 168.5 171.8 159.0 155.3
12 137.9 138.9 125.9 115.2
Average 184.2 186.5 177.7 171.9
13 152.6 153.7 147.1 141.9
Overall average 167.0 168.3 160.2 155.0

from the performance of Group III instances, for which HM re-
quires 165 seconds on average. The combined effect of time win-
dows and load can be analyzed by looking at the average CPU
time of instances Group IV, which is about 171 seconds. Group
IV, namely the instance set of thirteen, is solved in 142 seconds.
Computational results of the performance measurements are
summarized in Table 5. This table presents, for each method, the
number of Pareto solutions found, as well as the values of the
hypervolume and the -indicators. These results confirm that HM
performs very well for all methods, with an average number of
40 Pareto solutions across all instances. The ECM yields the second
best performance after HM. The number of Pareto solutions found
are similar for WM and WMN, which are 11.6 and 9.4, respectively.
We now compare the four methods using the two quality indica-
tors. According to the hypervolume indicator, HM is superior to
the other methods with an average value of 1386.3. In other words,
the solutions found by HM represents a larger area than the other
methods. The same indicator yields very poor results for WM and
WMN. With the -indicator, HM once again exhibits the best per-
formance, yielding the minimum value. The WM and ECM have a
very similar performances based on -indicator, whereas WMN
performs the worst.
We have also looked the effect of increasing a number of vehi-
cles on the performance measure. For each method, Table 6 pre-
sents the number of Pareto optimal solutions found, the
hypervolume and -indicators. In this table, the instance sets are
grouped by the number of vehicles to see the effect on the resulting
solutions.
The analysis of Table 6 shows that the average number of Pareto
solutions increases with the number of vehicles. This is because the
solution space is enlarged when more vehicles are needed. Finally,
the number of successes of each method using the two indicators is
reported in Table 7.
All results show that WM, WMN and ECM are clearly dominated
by HM, as far as the success rate is concerned, indicating that a hy-
brid use of the existing methods is better suited to our problem.
The weighting methods (WM and WMN) are better for finding ex-
treme (corner) solutions, but they do not generate many Pareto
solutions. On the other hand, the -constraint method finds more
solutions but most of them are inferior to those provided by WM.

Table 5
Results of quality indicators on bi-objective PRP instances.

Instance WM WMN ECM HM

sets
# Of Pareto Ihv ðSÞ I ðS; RÞ # Of Pareto Ihv ðSÞ I ðS; RÞ # Of Pareto Ihv ðSÞ I ðS; RÞ # Of Pareto Ihv ðSÞ I ðS; RÞ
solutions solutions solutions solutions
1 3.5 21.44 1.1276 2.3 19.21 1.1264 17.4 117.22 1.1255 25.2 118.46 1.1271
5 5.7 67.52 1.1058 3.4 67.96 1.1056 33.8 302.98 1.1054 29.1 334.25 1.1071
9 6.6 440.08 1.1326 3.2 460.55 1.1334 48.9 3958.28 1.1330 60.5 4004.49 1.1340
2 12.3 44.47 1.1421 11.8 43.03 1.1484 19.7 61.97 1.1294 21.1 71.11 1.1227
6 18.1 359.63 1.1283 17.9 388.13 1.1267 28.5 711.99 1.1235 39.4 746.67 1.1081
10 25.0 1519.07 1.1324 19.7 1478.08 1.1407 49.4 4102.16 1.1413 62.7 4266.04 1.1225
3 3.7 40.52 1.1198 3.6 39.98 1.1196 26.4 156.81 1.1207 30.0 164.63 1.1250
7 4.4 70.92 1.1037 3.3 82.20 1.1031 26.0 331.50 1.1033 31.2 345.82 1.1019
11 4.7 441.61 1.1519 2.7 369.18 1.1532 42.2 3084.67 1.1551 65.7 3343.63 1.1541
4 11.5 94.72 1.1330 10.0 82.88 1.1412 23.2 194.40 1.1312 31.0 199.08 1.1285
8 18.6 165.01 1.1317 14.9 142.35 1.1314 28.5 208.28 1.1292 34.8 289.00 1.1203
12 20.2 1280.08 1.1483 14.5 1319.75 1.1499 48.6 3266.90 1.1481 58.6 3821.56 1.1299
13 16.6 176.57 1.1231 14.7 162.08 1.1304 25.7 293.10 1.1289 30.0 317.12 1.1131

Average 11.6 363.20 1.1293 9.4 358.11 1.1315 32.2 1291.56 1.1288 39.9 1386.30 1.1226
 E. Demir et al. / European Journal of Operational Research 232 (2014) 464–478 473

Table 6
Results of quality indicators on benchmark instances grouped by the number of vehicles.

Instance WM WMN ECM HM

sets
# Of Pareto Ihv ðSÞ I ðS; RÞ # Of Pareto Ihv ðSÞ I ðS; RÞ # Of Pareto Ihv ðSÞ I ðS; RÞ # Of Pareto Ihv ðSÞ I ðS; RÞ
solutions solutions solutions solutions
1–4 7.8 50.29 1.1306 6.9 46.27 1.1339 21.7 132.60 1.1267 26.8 138.32 1.1258
5–8 11.7 165.77 1.1174 9.9 170.16 1.1167 29.2 388.69 1.1153 33.6 428.93 1.1094
9–12 14.1 920.21 1.1413 10.0 906.89 1.1443 47.3 3603.00 1.1444 61.9 3858.93 1.1351
13 16.6 176.57 1.1231 14.7 162.08 1.1304 25.7 293.10 1.1289 30.0 317.12 1.1131
Average 12.5 328.21 1.1281 10.4 321.35 1.1313 31.0 1104.35 1.1288 38.1 1185.83 1.1208

5.5. Details of the Pareto solutions
In this part of the analysis, we look at the nature of the Pareto
solutions identified by the algorithm for the speed level between
20 and 100 kilometer/hour. For this purpose, we provide graphs
of a sample instance for comparing the four methods in Fig. 3,
and detailed results in Table 8. In Fig. 3, the values on the x-axes
represent the driving time objective (f2) and the values on the y-
axes show the fuel consumption (f1).
Fig. 3(a)–(b) exhibit the Pareto solutions found by each method.
A recurring theme in the results presented in Fig. 3(d) is that driv-
ing time can be decreased from about 39 hours to 37 hours,
depending on the instance tested, without much change in the fuel
consumption. Conversely, fuel consumption can be brought down
quite significantly, from around 305 liters to 285 liters with only
a slight increase in driving time. Furthermore, all four methods
tested here are consistent with respect to the solutions generated
and yield similar insights. Interestingly, these results on the
trade-off between fuel consumption and driving time show that
one does not necessarily need to sacrifice one objective heavily
in order to improve the other. The tools presented in this paper of-
fer ways in which good quality solutions for the two objectives can
be found.
Table 8 shows detailed results of this experiment found by HM
for all 13 instance sets, in particular the minimum and maximum
values for cost, time and fuel consumption for each instance.
Table 8 shows that the results heavily depend on the number of
vehicles used. For group III, for example, savings in fuel can be up
to 200 liters for around 20 vehicles, and the driving time can also
be decreased by up to nine hours, depending on the instance
tested. In other words, the total fuel consumption and driving time
increase when more routes are needed. This means that more sav-
ings in fuel or driving time can be achieved. Also, for instance set
12, one achieves a reduction on fuel consumption of about 52%
by increasing the driving time by about 19%.
5.5.1. Experiments with practical speed limits
In this section, we look at the nature of the Pareto solutions
identified by the algorithm for more practical lower and upper
speed levels, set to 60 kilometer/hour and 112 kilometer/hour,
respectively (The Government of the United Kingdom, 2013). Using
the same instances as for Fig. 3, we provide similar graphs with the
settings obtained by using the four methods. The new results are
presented in Fig. 4. Detailed results for 13 instances are provided
in Table 9 in the same way as in Table 8.
The results reported in Table 9 do not significantly differ from
those of Table 8. In particular, the average differences for cost,
fuel and time are £99.9, 82.9 liters and 4.1 hours, respectively,
when the speed limits lie between 60 kilometer/hour and
112 kilometer/hour. The same statistics are observed for the case
where speed limits are chosen between 20 kilometer/hour and
100 kilometer/hour are £92.7, 81.1 liters and 3.9 hours,
respectively.
We have also calculated some statistics on the resulting
speeds. When speed limits lie between 20 kilometer/hour and
100 kilometer/hour, the minimum speed used in any solution
is 55 kilometer/hour, which is the speed minimizing fuel con-
sumption, as shown in Fig. 1. For the same case, 85.4% of all
speeds obtained across the instances are contained within the
interval [70, 85), whereas the remaining 8.9% lie the interval
[85, 100]. When speed limits of 60 kilometer/hour and 112 kilo-
meter/hour are enforced, 5.6% of all speeds are contained within

Table 7
Number of successes based on each instance set.

Instance sets WM WMN ECM HM

Ihv ðSÞ I ðS; RÞ Ihv ðSÞ I ðS; RÞ Ihv ðSÞ I ðS; RÞ Ihv ðSÞ I ðS; RÞ
1 0 2 0 4 5 2 5 2
2 0 2 0 0 3 3 7 5
3 0 2 1 4 3 3 6 1
4 0 1 0 2 3 3 7 4
5 1 2 0 4 4 3 5 1
6 0 0 0 1 4 2 6 7
7 0 1 0 2 3 4 7 3
8 1 2 0 1 3 2 6 5
9 0 2 0 3 7 4 3 1
10 0 2 0 1 6 0 4 7
11 0 5 0 3 2 1 8 1
12 0 1 0 0 3 0 7 9
13 0 1 0 2 1 1 9 6

Total 2 23 1 27 47 28 80 52
474 E. Demir et al. / European Journal of Operational Research 232 (2014) 464–478

Fig. 3. Pareto optimal solutions found for a sample instance.

Table 8
The found extreme Pareto solutions from each set with HM.

Instance sets Cost (£) Difference (£) Fuel (liter) Difference (liter) Time (hour) Difference (hour)
Min. Max. Min. Max. Min. Max.
1 451.0 472.8 21.8 193.9 216.6 22.7 20.6 22.8 2.2
2 463.0 479.4 16.4 194.8 221.4 26.6 21.2 24.1 2.9
3 454.9 495.3 40.4 195.1 234.8 39.7 20.8 23.0 2.2
4 463.2 483.4 20.1 195.1 218.8 23.6 21.2 23.9 2.7

5 618.0 678.7 60.0 243.3 301.1 57.8 30.9 34.7 3.7

6 691.7 722.5 30.8 273.0 305.1 32.1 36.9 39.1 2.1
7 613.2 640.4 27.1 240.7 273.9 33.2 32.1 34.5 2.4
8 646.2 675.7 29.6 249.9 286.9 37.0 33.4 37.7 4.3

9 910.3 1119.1 208.8 356.5 568.5 212.0 46.8 51.4 4.6

10 920.1 1146.0 225.9 357.0 545.6 188.6 47.8 53.3 5.6
11 925.5 1184.7 259.2 363.9 506.6 142.8 46.3 52.0 5.8
12 933.1 1169.4 236.2 361.9 559.8 197.9 45.5 54.3 8.8

13 691.5 720.4 28.9 275.4 315.7 40.4 34.5 38.5 4.0

the interval [60, 75), 86.1% lie within the interval [75, 90), and
the rest are in the interval [90, 112]. This simple analysis shows
that the resulting speeds are not very dependent on the initial
range of speeds, and that the resulting solutions use reasonably
practical speeds. These experiments were performed with a un-
ique maximum speed for the entire network. We are aware that
different speed limits may apply to various route segments, but
the analysis we have performed enables us to derive meaningful
conclusions on the nature of Pareto solutions.
5.6. Results for a sample 30-node instance
This section presents the results of a 30-node instance. Our aim
is to take a closer look at the trade-off between the two objectives.
Table 10 presents two non-dominated solutions obtained by HM
on the total fuel consumption (in liter), the total driving time (in
hour), the total fuel cost, the total driving time and the total
amount of CO2. In calculating the latter, we assume that a liter of
diesel fuel produces 3.135 kilograms of CO2e.
 E. Demir et al. / European Journal of Operational Research 232 (2014) 464–478 475

Fig. 4. Pareto optimal solutions found for a sample instance with realistic speed levels.

Table 9
The found extreme Pareto solutions from each set with HM.

Instance sets Cost (£) Difference (£) Fuel (liter) Difference (liter) Time (hour) Difference (hour)
Min. Max. Min. Max. Min. Max.
1 450.2 493.7 43.5 194.6 224.1 29.5 20.1 22.7 2.6
2 459.2 481.7 30.5 195.2 221.3 26.1 21.2 23.2 2.0
3 451.2 483.7 32.5 196.2 219.7 23.5 20.9 22.8 1.9
4 466.2 489.3 23.1 198.1 222.3 24.2 20.6 23.9 3.3

5 617.2 703.4 86.2 252.3 328.9 66.6 30.2 33.8 3.6

6 701.2 734.2 33.0 281.1 318.2 37.1 35.4 38.1 2.7
7 623.5 678.2 54.7 244.2 289.1 44.9 31.9 34.1 2.3
8 656.2 678.1 21.9 250.2 297.9 47.7 33.1 37.2 4.1

9 930.2 1141.7 211.5 362.3 584.7 210.4 45.6 50.2 4.6

10 937.8 1174.9 237.1 363.7 557.6 193.9 46.8 52.8 6.0
11 929.9 1201.3 271.4 367.1 513.2 146.1 45.2 50.9 5.7
12 947.1 1178.2 231.1 371.2 564.4 193.2 44.1 53.6 9.5

13 709.8 732.5 22.7 281.9 319.1 37.2 33.6 38.2 4.6

Solution A and B are depicted in Fig. 5, with each route shown in
a different shade. Solution A (Fig. 5(a)) shows a time-minimizing
tour of length 1621.7 kilometers consuming approximately
321.57 liters diesel fuel. The tour needs 21.16 hours to be tra-
versed, and the total amount of CO2e emitted is 1008.12 kilograms.
Solution B (Fig. 5(b)) shows a fuel-minimizing tour of length
1270.1 kilometers consuming around 233.54 liters diesel fuel.
The total amount of CO2e for this tour is around 732.15 kilograms,
with a driving time of 23.21 hours and a total distance of
1270.1 kilometers.
476 E. Demir et al. / European Journal of Operational Research 232 (2014) 464–478
Table 10
Two non-dominated solutions of 30-node instance.
# Of Total distance Fuelconsumption
routes (kilometer) (liter)
Solution 6 1621.7 321.57
A
Solution 6 1270.1 233.54
B
Fig. 5. Two non-dominated solutions for a 30-node instance.
The trade-off between fuel consumption and time for this par-
ticular instance indicates that savings in energy can be achieved
by increasing the total duration of routes. The 9.7% increase in driv-
ing time leads to a 27% saving in energy requirements. The reduc-
tion of CO2e is around 276 kg if Solution B is preferred to Solution
A. In contrast, the reduction in driving time from 23.21 hours to
21.16 hours (about 8.8%) afforded by solution A implies an increase
in CO2e emissions of about 37.7%.
6. Conclusions
We have studied the bi-criteria PRP in which one of the objec-
tives is related to CO2e emissions, and the other to driving time.
An enhanced adaptive large neighborhood search (ALNS) algorithm
was used for the generation of non-dominated/Pareto optimal
solutions. The algorithm integrates the classical ALNS with a spe-
cialized speed optimization procedure. The proposed algorithm
first calls the ALNS using fixed speeds as inputs, then optimizes
speed on each route. Our approach implicitly assumes that vehicle
speed can be optimized independently of the remaining road traf-
fic. Taking congestion into account cannot readily be incorporated
in a model such as ours since full traffic data are not available at
the planning stage. This is clearly a limitation of our study.
Using the ALNS as the search engine, four a posteriori methods,
namely the weighting method, the weighting method with nor-
malization, the epsilon-constraint method and a new hybrid meth-
od (HM), were tested using a scalarization of the two objective
functions. The HM combines an adaptive weighting with the epsi-
lon-constraint method. To fully evaluate the effectiveness of the
algorithm, new sets of instances based on real geographic data
Operational time CO2e emissions Fuel cost Driver cost Total cost
(hour) (kilogram) (£) (£) (£)
21.16 1008.12 450.20 169.28 619.48
23.21 732.15 326.96 185.68 512.64
were generated, and a library of bi-criteria PRP instances was com-
piled. Results of extensive computational experimentation of the
four methods were presented and compared with one another.
Hypervolume and epsilon indicators were used to evaluate the per-
formance of the four methods. Our results show that the HM is
highly effective in finding good-quality non-dominated solutions
on 100-node instances, both in terms of the hypervolume and of
the epsilon indicators, as compared to the rest of the methods.
The benefits of tools described in the paper can be summarized
as follows: (i) to establish the cost of reducing CO2e emissions
through routing decisions; (ii) to help achieve CO2e neutral trans-
portation and use speed adaptation as one of the ways of achieving
it; and (iii) to be prepared for changes (i.e. increases) in fuel prices
without having to reschedule the routes over and over again, i.e.,
create a ‘‘robust’’ schedule.
One interesting insight derived from the experiments con-
ducted here is that there is no need to greatly compromise on driv-
ing time in order to achieve a significant reduction in fuel
consumption and CO2e emissions. The converse of this conclusion
also holds, at least for the instances tested in this paper, i.e., consid-
erable reductions in driving time may be obtained if one is willing
to increase fuel consumption only slightly. Our results indicate that
the trade-offs between the conflicting objectives of fuel and time
do not necessarily require great sacrifices to be made with respect
to an objective in order to improve the other. However, we urge
caution in that these results may not apply to the entire logistics
industry. One particular example arises in systems or industries
employing a just-in-time policy, where the cost functions used
may be such that the cost of late deliveries easily outweighs the
CO2e costs.
 E. Demir et al. / European Journal of Operational Research 232 (2014) 464–478
Other costs, such as equipment and labor which are not consid-
ered in this paper, could also affect our conclusions on the time vs.
fuel objective, provided that these costs are adequately reflected
within the model presented in this paper. The tools described here
should help decision makers by providing them with a set of solu-
tions from which they can choose.
One immediate use of the approach described here is when
managers or route-planners are faced with carbon taxes in the
event that they exceed prescribed amounts of emissions. Such
taxes are already being imposed in the airline industry. Through
the use of the tool proposed in this paper, managers would then
be able to assess the impact of exceeding tolerable levels of emis-
sion and weigh the effects of taxes against other measures, such as
time and prompt delivery. Again, the fact that our approach is
independent of costs implies the use of the technique for differing
values of carbon tax, which is likely to be the case from one coun-
try or geographical region to another.
Acknowledgements
The authors gratefully acknowledge funding provided by the
University of Southampton School of Management and by the
Canadian Natural Sciences and Engineering Research Council under
Grant 39682-10. Thanks are due to two anonymous reviewers for
their useful comments and for raising interesting points for
discussion.
References
Apaydin, O., & Gonullu, M. T. (2008). Emission control with route optimization in
solid waste collection process: A case study. Sadhana, 33(2), 71–82.
Barth, M., & Boriboonsomsin, K. (2008). Real-world CO2 impacts of traffic
congestion. Transportation Research Record: Journal of the Transportation
Research Board, 2058(1), 163–171.
Barth, M., Younglove, T., & Scora, G. (2005). Development of a heavy-duty diesel
modal emissions and fuel consumption model. Technical report, UC Berkeley:
California Partners for Advanced Transit and Highways (PATH).
Bektasß, T., & Laporte, G. (2011). The pollution-routing problem. Transportation
Research Part B: Methodological, 45(8), 1232–1250.
Bowyer, D. P., Biggs, D. C., & Akçelik, R. (1985). Guide to fuel consumption analysis
for urban traffic management. Australian Road Research Board Transport
Research, 32.
Chankong, V., & Haimes, Y. Y. (1983). Optimization-based methods for
multiobjective decision-making: An overview. Large Scale Systems, 5(1), 1–33.
Clarke, G., & Wright, J. W. (1964). Scheduling of vehicles from a central depot to a
number of delivery points. Operations Research, 12(4), 568–581.
Coello, C. A. C., Lamont, G. B., & Van Veldhuizen, D. A. (2002). Evolutionary algorithms
for solving multi-objective problems. Dordrecht: Kluwer.
Crainic, T. G. (2000). Service network design in freight transportation. European
Journal of Operational Research, 122(2), 272–288.
Deb, K. (2001). Multi-objective optimization using evolutionary algorithms. Chichester,
UK: Wiley.
Deb, K., & Chaudhuri, S. (2007). I-mode: an interactive multi-objective optimization
and decision-making using evolutionary methods. Lecture Notes in Computer
Science, 4403, 788–802.
Demir, E., Bektasß, T., & Laporte, G. (2011). A comparative analysis of several vehicle
emission models for road freight transportation. Transportation Research Part D:
Transport and Environment, 6(5), 347–357.
Demir, E., Bektasß, T., & Laporte, G. (2012). An adaptive large neighborhood search
heuristic for the pollution-routing problem. European Journal of Operational
Research, 223(2), 346–359.
Eglese, R. W., Maden, W., & Slater, A. (2006). A road timetable™ to aid vehicle
routing and scheduling. Computers and Operations Research, 33(12), 3508–3519.
Ehrgott, M., & Gandibleux, X. (2002). Multiple criteria optimization: State of the art
annotated bibliographic surveys. Dordrecht: Kluwer.
Ehrgott, M., & Ryan, D. M. (2002). Constructing robust crew schedules with
bicriteria optimization. Journal of Multi-Criteria Decision Analysis, 11(3),
139–150.
Figliozzi, M. A. (2010). Vehicle routing problem for emissions minimization.
Transportation Research Record: Journal of the Transportation Research Board,
2197(-1), 1–7.
Figliozzi, M. A. (2011). The impacts of congestion on time-definitive urban freight
distribution networks co2 emission levels: Results from a case study in Portland,
Oregon. Transportation Research Part C: Emerging Technologies, 19, 766–778.
Fonseca, C. M., & Fleming, P. J. (1995). An overview of evolutionary algorithms in
multiobjective optimization. Evolutionary Computation, 3(1), 1–16.
477
Forkenbrock, D. J. (1999). External costs of intercity truck freight transportation.
Transportation Research Part A: Policy and Practice, 33(7–8), 505–526.
Forkenbrock, D. J. (2001). Comparison of external costs of rail and truck freight
transportation. Transportation Research Part A, 35(4), 321–337.
Fowkes, T., & Whiteing, T. (2006). The value of freight travel time savings and
reliability improvements – Recent evidence from Great Britain. In European
Transport Conference (ETC). Strasbourg, France: Association for European
Transport.
Grodzevich, O., & Romanko, O. (2006). Normalization and other topics in multi-
objective optimization. In Proceedings of the Fields MITACS Industrial Problems
Workshop, Toronto.
Hickman, J., Hassel, D., Joumard, R., Samaras, Z., & Sorenson, S. (1999). MEET-
methodology for calculating transport emissions and energy consumption.
European Commission, DG VII. Technical report, ISBN 92-828-6785-4,
Luxembourg, 362 p. <http://www.transport-research.info/Upload/Documents/
200310/meet.pdf> Accessed 05.05.13.
Jabali, O., Van Woensel, T., & de Kok, A. G. (2012). Analysis of travel times and CO2
emissions in time-dependent vehicle routing. Production and Operations
Management, 21(6), 1060–1074.
Jozefowiez, N., Semet, F., & Talbi, E. G. (2008a). From single-objective to multi-
objective vehicle routing problems: Motivations, case studies, and methods. In
B. L. Golden, S. Raghavan, & E. A. Wasil (Eds.), The vehicle routing problem: Latest
advances and new challenges (pp. 445–471). New York: Springer.
Jozefowiez, N., Semet, F., & Talbi, E. G. (2008b). Multi-objective vehicle routing
problems. European Journal of Operational Research, 189(2), 293–309.
Kirkpatrick, S., Gelatt, C. D., Jr., & Vecchi, M. P. (1983). Optimization by simulated
annealing. Science, 220(4598), 671–680.
Laumanns, M., Thiele, L., & Zitzler, E. (2006). An efficient, adaptive parameter
variation scheme for metaheuristics based on the epsilon-constraint method.
European Journal of Operational Research, 169(3), 932–942.
Maden, W., Eglese, R. W., & Black, D. (2009). Vehicle routing and scheduling with
time-varying data: A case study. Journal of the Operational Research Society,
61(3), 515–522.
Mavrotas, G. (2009). Effective implementation of the [epsilon]-constraint method in
multi-objective mathematical programming problems. Applied Mathematics and
Computation, 213(2), 455–465.
McKinnon, A. (2007). CO2 Emissions from Freight Transport in the UK. Report
prepared for the Climate Change Working Group of the Commission for
Integrated Transport.
Miettinen, K. (1999). Nonlinear multiobjective optimization. Norwell: Kluwer.
Ntziachristos, L., & Samaras, Z. (2000). COPERT III computer programme to calculate
emissions from road transport: methodology and emission factors (version 2.1).
Technical report, European Environment Agency, Copenhagen, Denmark.
<http://lat.eng.auth.gr/copert/files/tech50.pdf> Accessed 05.05.13.
Palmer, A. (2007). The development of an integrated routing and carbon dioxide
emissions model for goods vehicles. PhD thesis, Cranfield University, Bedford,
United Kingdom.
Paquette, J., Cordeau, J.-F., Laporte, G., & Pascoal, M. M. B. (2013). Combining
multicriteria analysis and tabu search for dial-a-ride problems. Transportation
Research Part B: Methodological, 52, 1–16.
Pitera, K., Sandoval, F., & Goodchild, A. (2011). Evaluation of emissions reduction in
urban pickup systems. Transportation Research Record: Journal of the
Transportation Research Board, 2224(1), 8–16.
PSRC (2010). Chapter 6 air quality and climate change. Technical report, Puget
Sound Regional Council, Washington.
Qian, F., Gribkovskaia, I., Laporte, G., & Halskau, Ø. (2011). Passenger and pilot risk
minimization in offshore helicopter transportation. Omega, 40(5), 584–593.
Rangaiah, G. P. (2009). Multi-objective optimization: Techniques and applications in
chemical engineering (Vol. 1). Singapore: World Scientic.
Ropke, S., & Pisinger, D. (2006). An adaptive large neighborhood search heuristic for
the pickup and delivery problem with time windows. Transportation Science,
40(4), 455–472.
Scora, M., & Barth, G. (2006). Comprehensive modal emission model (CMEM),
version 3.01, user’s guide. Technical report.
Scott, C., Urquhart, N., & Hart, E. (2010). Influence of topology and payload on CO2
optimised vehicle routing. Applications of Evolutionary Computation, 141–150.
Shaw, P. (1998). Using constraint programming and local search methods to solve
vehicle routing problems. In Proceedings of the 4th International conference on
principles and practice of constraint programming (pp. 417–431). New York:
Springer.
Tavares, G., Zsigraiova, Z., Semiao, V., & Carvalho, M. G. (2009). Optimisation of
MSW collection routes for minimum fuel consumption using 3D GIS modelling.
Waste Management, 29(3), 1176–1185.
Tavares, G., Zsigraiova, Z., Semiao, V., & da Graça Carvalho, M. (2008). A case study of
fuel savings through optimisation of MSW transportation routes. Management
of Environmental Quality: An International Journal, 19(4), 444–454.
Tight, M. R., Bristow, A. L., Pridmore, A., & May, A. D. (2005). What is a sustainable
level of CO2 emissions from transport activity in the UK in 2050? Transport
Policy, 12(3), 235–244.
Ubeda, S., Arcelus, F. J., & Faulin, J. (2011). Green logistics at Eroski: A case study.
International Journal of Production Economics, 131(1), 44–51.
The Government of the United Kingdom (2013). Speed limits, May. <https://
www.gov.uk/speed-limits>.
Urquhart, N., Hart, E., & Scott, C. (2010a). Building low CO2 solutions to the vehicle
routing problem with Time Windows using an evolutionary algorithm. In
Evolutionary computation (CEC), 2010 IEEE congress (pp. 1–6). IEEE.
478 E. Demir et al. / European Journal of Operational Research 232 (2014) 464–478
Urquhart, N., Scott, C., & Hart, E. (2010b). Using an evolutionary algorithm to
discover low CO2 tours within a travelling salesman problem. Applications of
Evolutionary Computation, 421–430.
Veldhuizen, D. A. V., & Lamont, G. B. (2000). Multiobjective evolutionary algorithms:
Analyzing the state-of-the-art. Evolutionary Computation, 8(2), 125–147.
Vicente, A. S. (2011). Laying the foundations for greener transport. term 2011:
Transport indicators tracking progress towards environmental targets in
Europe. EEA Report 7/2011, European Environment Agency, Copenhagen,
Denmark.
Wierzbicki, A. P. (1999). Reference point approach. In Multicriteria decision making:
Advances in MCDM models, algorithms, theory, and applications (pp. 9–39).
Dordrecht: Kluwer.
Wright, L. A., Kemp, S., & Williams, I. (2011). Carbon footprinting: Towards a
universally accepted definition. Carbon, 2(1), 61–72.
Zhou, A., Qu, B. Y., Li, H., Zhao, S. Z., Suganthan, P. N., & Zhang, Q. (2011).
Multiobjective evolutionary algorithms: A survey of the state-of-the-art. Swarm
and Evolutionary Computation, 1(1), 32–49.
Zitzler, E., Deb, K., & Thiele, L. (2000). Comparison of multiobjective evolutionary
algorithms: Empirical results. Evolutionary Computation, 8(2), 173–195.
Zitzler, E., & Thiele, L. (1998a). An evolutionary algorithm for multiobjective
optimization: The strength pareto approach. Technical report, TIK Report,
Zurich.
Zitzler, E., & Thiele, L. (1998b). Multiobjective optimization using evolutionary
algorithms: A comparative case study. In 1998: Proceedings on parallel problem
solving from nature-PPSN V: 5th international conference, Amsterdam, The
Netherlands, September 27–30 (number 1498, p. 292). Amsterdam: Springer-
Verlag.
Zitzler, E., Thiele, L., Laumanns, M., Fonseca, C. M., & da Fonseca, V. G. (2003).
Performance assessment of multiobjective optimizers: An analysis and review.
IEEE Transactions on Evolutionary Computation, 7(2), 117–132.