Expert Systems with Applications 41 (2014) 4245–4258

Contents lists available at ScienceDirect

Expert Systems with Applications

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

Green logistic vehicle routing problem: Routing light delivery vehiclesin urban areas using a neuro-fuzzy model

0957-4174/$ - see front matter � 2014 Elsevier Ltd. All rights reserved.http://dx.doi.org/10.1016/j.eswa.2014.01.005

⇑ Corresponding author. Address: Pavla Jurisica Sturma 33, 11000 Belgrade,Serbia. Tel.: +381 642377908; fax: +381 113603187.

E-mail addresses: cirovic@sezampro.rs (G. Cirovic), dpamucar@gmail.com (D.Pamucar), dbozanic@yahoo.com (D. Bozanic).

1 Address: Hajduk Stankova 2, 11000 Belgrade, Serbia. Tel.: +381 112422178; fax:+381 112422178.

2 Address: Pavla Jurisica Sturma 33, 11000 Belgrade, Serbia. Tel.: +381 668700367;fax: +381 113603187.

Goran Cirovic a,1, Dragan Pamucar b,⇑, Darko Bozanic c,2

a The Belgrade University College of Civil Engineering and Geodesy, Serbiab University of Defence in Belgrade, Department of Logistic, Serbiac University of Defence in Belgrade, Military Academy, Serbia

a r t i c l e i n f o

Keywords:Environmentally friendly vehiclesVehicle routingGreen logisticsNeuro-fuzzySimulated annealing

a b s t r a c t

Today’s growth in the level of traffic in cities is leading to both congestion and environmental pollution(exhaust emissions and noise), as well as increased costs. Traffic congestion makes cities less pleasantplaces to live in, a particular problem being the negative impact on health as a result of increased exhaustemissions. In addition to these emissions, another major effect of transport which can lead to serioushealth problems is noise (EEA, 2013a, 2013b). There is a strong tendency in the world towards the devel-opment of ‘‘clean’’ motor vehicles that do not pollute the environment, that is, that do not emit harmfulsubstances in their exhaust fumes and which create less noise without causing other types of pollution.The growth in the influence of transport on the environment has resulted in planners formulating proce-dures which take into account the effect of traffic on the quality of life in urban areas. This paper presentsa model for the routing of light delivery vehicles by logistics operators. The model presented takes intoaccount the fact that logistics operators have a limited number of environmentally friendly vehicles (EFV)available to them. When defining a route, EFV vehicles and environmentally unfriendly vehicles (EUV) areconsidered separately. For solving the problem of routing in the model, an adaptive neural network wasused which was trained by a simulated annealing algorithm. An adaptive neural network was used forassessing the performance of the network branches. The input parameters of the neural network werethe logistics operating costs and environmental parameters (exhaust emissions and noise) for the givenvehicle route. Each of the input parameters of the neural network was thoroughly examined. The inputparameters were broken down into elements which further describe the state of the environment, noiseand logistics operating costs. After obtaining the performance of the network links for calculating theroute for EFV and EUV vehicles a modified Clark–Wright algorithm was used. The proposed model wastested on a network which simulates the conditions in the very centre of Belgrade. All of the input param-eters of the model were obtained on the basis of 40 automatic measuring stations for monitoring the airquality (SEA, 2012).

� 2014 Elsevier Ltd. All rights reserved.

1. Introduction

In its simplest possible form, logistics can be described as the pro-cess of delivering a product (service) in the required quantities, ingood condition, to the appointed place at the appointed time, to aspecific customer at an agreed price. The expansion of logistics camealong with the growing trend of globalization and decentralization

of production, the functioning of which depends significantly onthe quality of logistics activities. The area of logistics has constantlyexpanded and developed, adapting to the demands of technologyand the environment. Today, logistics is present in all areas of soci-ety. At the same time as its development, there has been a growth inenvironmental awareness, and it is indeed noticeable that in the21st century environmental problems occupy an important placein the priority list of the world’s problems.

Today it is difficult to imagine any system without logisticssupport. However, the realization of key logistics processes(transport, handling, storage) conflicts with the requirements forenvironmental protection, with transport being characterized asone of the major environmental pollutants. For precisely thisreason, this paper focuses on the organization of the transport pro-cess, more specifically on green transport within the framework of

4246 G. Cirovic et al. / Expert Systems with Applications 41 (2014) 4245–4258

green logistics, which during the implementation of the logisticsprocess, uses an approach involving environmental preservation.

According to research carried out in Japan and Great Britain(Murphy & Poist, 2003), heavy goods vehicles generate noise be-tween 88 and 92 dB, and light goods vehicles between 79 and81 dB. If we bear in mind that noise which exceeds 60 dB isharmful to human health, then it is not difficult to conclude thaton this basis road transport is the greatest cause of adverse ef-fects on the environment. Traffic noise, as the main source ofnoise in urban areas, is a very significant ecological problem interms of its serious damage to the health of the population, alsocausing a reduction in labor productivity. Recent data on thethreat of environmental noise to the population was obtainedafter the first round in the production of strategic noise mapsfor agglomerations in European Union countries. Data from theEuropean Environment Agency (2013a, 2013b) indicate that in ur-ban areas 54% of the population (56,001,200 people) are exposedto full-day noise levels exceeding 55 dB(A) and 15% of the popu-lation (15,754,500 individuals) full-day noise levels greater than65 dB(A). In addition, away from the agglomeration another33,437,244 residents live in areas where the full-day noise levelis greater than 55 dB(A), of whom 7,657,083 residents are in areaswhere the full-day noise level is greater than 65 dB(A). From a to-tal of 89,438,444 residents who are subject to full-day levels ofnoise greater than 55 dB(A) almost 89 million are exposed tonoise generated by traffic (road, rail and air). The number of peo-ple exposed to full-day levels of noise greater than 55 dB(A) aris-ing from road traffic is almost 68 million, indicating road traffic tobe the dominant source of noise. A report by the European Envi-ronment Agency (2013a, 2013b) showed that in most Europeancities, three out of five people are exposed to harmful levels oftraffic noise. Research carried out by Schreyer et al. (2004) hasshown that in the European Union the external costs of trafficaccidents occurring as a result of the deterioration of air qualityand an increase in noise level amounts to between 0.5% and3.7% of the gross domestic product of EU countries.

Many industrialized countries and developing countries haveadopted different regulations to define the maximum permissiblelevel of noise emissions of motor vehicles, the noise inside thevehicle as well as noise emissions in urban areas. However, evenwith very ‘‘quiet’’ motor vehicles and laws favoring their use, it willbe many years before existing outdated fleets are completely re-placed. Bearing this in mind, the model presented in this paper rec-ognises the fact that logistics operators in urban zones use not onlyEFV but also EUV vehicles.

Besides noise, another significant consequence of transportthat can lead to serious health problems is harmful exhaust emis-sions (EEA, 2013a; EEA, 2013b). Transport is responsible for about14% of the total carbon dioxide emissions (CO2) (Yazan, Petruzz-elli, & Albino, 2011). Of the total amount of harmful substancesemitted 50% is caused by traffic, while in urban areas its sharecan be as high as 90%. Road traffic, as a specific branch of traffic,significantly contributes to air pollution. This type of transportcauses 86% of the carbon monoxide (CO), 33% of the hydrocarbon(CH) and 42% of the nitrogen oxide (NOx) pollution. In addition,petrol engines are the main sources of lead pollution, and dieselengines emit large amounts of soot and smoke. Moreover, differ-ent types of transport have different rates of energy consumption(and thus different fuel emissions) for carrying out the sametransport task. The largest consumer is road transport, which ac-counts for 82% of the total energy consumption in transport, fol-lowed by air with 13%, rail with 3% and river with 2% (Bapna,Thakur, & Nair, 2002).

The increased demand placed on logistics distribution systemshas led to the construction of distribution centers and terminals

near or within urban areas. By developing the ‘‘just in time’’ con-cept, many production systems have almost lost their storage func-tion which not only required investment, but also generatedsignificant running costs. This has resulted in the transfer of certainstock to the transport system. Part of the stock is actually in transit,which causes greater congestion and pollution in cities, with theenvironment and society bearing the cost. This has been confirmedby empirical research in the UK, where out of a sample of 87 com-panies a 39% reduction in the number and capacity of storage facil-ities was recorded, while 1/3 of the companies recorded anincreased volume of delivery transport (McKinnon, 2008).

A large concentration of logistics activities in populated areascauses a great deal of air pollution as well as noise. One of thesolutions to this for logistics operators is the introduction ofEFV. These vehicles have found their place in the prevention ofglobal warming and reduction of pollution caused by CO,CO2, CH, NOx, SO2 (sulfur dioxide) and particulate emissions(PM10 and PM2.5). In addition, EFV are especially important interms of reducing the emission of vibrations and noise as specificforms of environmental pollution which are especially prevalentin urban areas. EFV vehicles are characterized by reduced emis-sions of harmful substances and are fuelled by gas, liquefiedpetroleum gas, ethanol, methanol, biodiesel, hydrogen, hybridand electrical energy.

Based on the environmental directives promoted by the Euro-pean Union, logistics operators in cities have begun to renew theirfleets by introducing EFV. As the number of these vehicles is cur-rently limited, implementation needs to be carried out togetherwith the existing vehicles in such a way to maximize the impacton reducing pollution. The dramatic increase in the impact of traf-fic on air quality in cities has influenced the emergence of new pro-cedures that will take into account the impact of transport on thequality of the environment in urban areas. There is no single solu-tion for all urban problems, but city authorities insist that logisticsoperators focus on an integrated approach in order to respond inthe best possible way to the problems that arise. This is achievedby combining knowledge from various areas of technology suchas the development of new vehicles, economic incentives andnew procedures for the creation of green EFV routes. In this paper,the term EFV of logistics operators is understood as light deliveryvehicles with reduced emission levels.

Several ecologically oriented extensions of the VRP have beenintroduced which aim at minimizing fuel consumption or theamount of CO2 emissions. For any of these problems, the evalua-tion of transportation plans relies on an estimation of the quan-tity of fuel consumed while completing the required task. Thereare a variety of methods for estimating the fuel consumptionand emissions of road transportation which depend on a wholerange of parameters. For an overview on methods see for exampleFrey, Zhang, and Rouphail (2010). Most of the estimation methodsare based on analytical emission models, and they differ in termsof the principles on which they are based and the parametersthey take into account for estimation. A comparison of severalvehicle emission models for road freight transportation can befound in Demir, Bektas, and Laporte (2011). In addition to com-paring different methods for estimating fuel consumption andpollution, Demir et al. (2011) analyze the discrepancies betweenthe results yielded by the models on the one hand and the resultsrecorded for the on-road consumption of real vehicles on theother hand.

Kara, Kara, and Yetis (2007) present a model for the problemthat minimizes the load weight carried by the vehicles. They claimthat their model aims at minimizing the energy required for therouted vehicles. More recent models are based on methods for esti-mating fuel and pollution which depend on specific parameters

G. Cirovic et al. / Expert Systems with Applications 41 (2014) 4245–4258 4247

and actually consider the total weight including the dead weight ofthe carrying vehicles, e.g. Peng and Wang (2009). These recentmodels take several factors into account, e.g. the average speed(Figliozzi, 2010), congestion influencing the average speed com-bined with acceleration rates (Figliozzi, 2010), topology (Scott,Urquhart, & Hart, 2010; Ubeda, Arcelus, & Faulin, 2010) and thepayload (Jaramillo, 2010; Peng & Wang, 2009; Scott et al., 2010).

An overview on issues linking Green Logistics with vehicle rout-ing and scheduling can be found in Sbihi and Eglese (2007) and inSbihi and Eglese (2010). In these papers the authors focus on as-pects of time-dependent problems, the transportation of hazardousmaterial and the dynamic optimization of real-time models.Related to these aspects, they discuss environmental objectives,as well as the characteristics of vehicle routing problems that in-volve the consideration of additional constraints. Jabali, Van Woen-sel, and de Kok (2012) study the trade-off between minimizing CO2

emissions and minimizing total travel times. As CO2 emissions aredirectly related to vehicle speed, time dependent travel times areincluded in their optimization models. Three different models arepresented and compared: a model for the minimization of the totaltravel time, a model for minimizing the total CO2 emission depend-ing on travel times and speed, and a cost-based model that opti-mizes a weighted average of travel time, emission and fuel costs.

The so-called Green Single Vehicle Routing Problem introducedin the articles by Jaramillo (2010) as well as Peng and Wang (2009)aims to minimize the the number of vehicles required for a roundtrip measured in the total ton miles related to the trip. Kuo (2010)also proposes a model for minimizing the total fuel consumptionfor the time-dependent vehicle routing problem where fuel con-sumption depends on speed, time of travel, and loading weight.The author presents a simulated annealing algorithm for solvingthis problem. In Kuo and Wang (2011) the problem which was pre-viously presented in Kuo (2010) is considered once more and thistime it is solved with a Tabu Search Algorithm instead of using asimulated annealing approach. Bektas and Laporte (2011) presentand compare several ecologically oriented extensions of the classi-cal VRP. These extensions are based on objective functions that ac-count not just for travel distance, but also for the amount ofgreenhouse emissions, fuel, travel times and costs. In their paper,mathematical models are described for these extended problemswith different orientations, such as distance-minimizing, weightedload-minimizing, energy-minimizing, and cost-minimizing.

Ubeda et al. (2010) analyze the effects of various degrees ofvehicle utilization on carbon dioxide emission. The authors analyzethe differences in CO2 emissions between a distance and an emis-sion minimization approach. Scott et al. (2010) investigate theinfluence of gradient and payload correction factors used withinCO2 emission models. The authors test the degree of influence onthe solutions of shortest path problems and traveling salesmanproblems when applied to freight delivery. For the estimation ofthe fuel consumption, their study employs the COPERT model pre-sented in Ntziachristos and Samaras (2000).

In addition to the research mentioned there have been a num-ber of studies published that discuss the organization of greentransportation and identify the problem of pollution. Bektas andLaporte (2011) developed a model for green vehicle routing whichis an extension of the classical VRP. In the developed model theobjective function is obtained by observing not only the Euclideandistance between nodes on the network and the travel time, butalso on the basis of the concentration of exhaust gases on the linkand fuel consumption. Xiao, Zhao, Kaku, and Xu (2011) pay specialattention to the rate of fuel consumption in relation to the vehicleload. Xiao et al. (2011) carried out optimization of the classic VRPproblem using a simulated annealing algorithm. The results showthat the model can reduce fuel consumption by an average of 5%compared to the standard VRP models. Erdogan & Miller-Hooks

(2012) introduce the Green Vehicle Routing Problem (G-VRP). Intheir paper they present the problem of routing a fleet of vehicleswhich uses biofuels and their modification of the Clark Wrightalgorithm using heuristic procedures. The use of vehicles whichrun on biofuels is considered when attempting to overcome thedifficulties which exist when using vehicles running on fossil fuels.The basic problems of using vehicles which run on fossil fuels areidentified as increased exhaust emissions (NOx, SO2 i CO) and alimited range of vehicles in combination with a limited infrastruc-ture for filling the vehicle with the required fuel.

In recent years, both the evolution of operational research andincrease in computing strength have prompted great interest inthis problem. Cipriani, Fusco, and Petrelli (2006) formulated thegreen vehicle routing problem as a problem of non-linear optimi-zation taking into account both discrete and constant variables. Agroup of Italian authors (Beltran, Carrese, Cipriani, & Petrelli,2009) investigated the implementation of green fleets of buses inan urban network, where they took into account all of the trafficin the network. Assuming user equilibrium in the network, theydeveloped a model for the allocation of the green fleets using ge-netic algorithms. They took sensitive areas into account (residen-tial zones, parks, etc.) and made a green route beside them,which had a great impact on the reduction of pollution.

Some of the most significant works in the area of EFV routing arethose by Baaj and Mahmassani (1995), Ceder and Israeli (1993) andCarrese and Gori (2002). They developed a new approach based onmetaheuristic techniques: genetic algorithms, simulated annealingand tabu search. In addition to the mentioned works, Ngamchai andLovell (2003) proposed the use of genetic algorithms for solving theproblem of EFV routing, while Fan and Machemehl (2006) consid-ered the transport requirements in terms of EFV distributionaccording to different types of transport. In addition to thesestudies, the problem of green vehicle routing has been dealt withby other authors (Faulin, Juan, Lera, & Grasman, 2011; Suzuki,2011). A detailed presentation and description of studies whichconsider the problem of vehicle routing, as well as the problem ofrouting ‘‘green’’ vehicles can be found in Lin, Choy, Ho, Chung,and Lam (2014). Examples of the application of a neuro-fuzzyapproach, and the theoretical basis for this approach can be foundin works by Shing and Jang (1993), Shing and Jang (1995) and Shiand Mizumoto (2000).

By analyzing the available literature, the conclusion can bereached that so far there has been no consideration of the problemof EFV routing using a neuro-fuzzy approach which takes into ac-count the paramaters of the environment and logistics operatingcosts and their impact on EFV routing. The problem lies in design-ing green EFV routes which minimize the harmful effects of trans-portation using light delivery vehicles in urban areas. Thisapproach becomes more important if we take into account the factthat in the future it can be expected that transport companies, par-ticularly in countries with developed industries, have an increasingnumber of EFVs in their fleets.

Local authorities are putting much effort into including as manyvehicles as possible with reduced emissions in their urban trans-port. However, there has been a noticeable lack of reliable method-ology to support such implementation. In order to optimize the‘‘green’’ capacity, a system has been developed to support deci-sion-making when routing light delivery vehicles with reducedemission levels in urban areas. The aim of this paper is to proposea model for the allocation of EFV in urban areas, taking into ac-count the environmental parameters on the given route. The prob-lem is presented as one of nonlinear optimization with fuzzyvalues of the input parameters, and is solved using neuro-fuzzy lo-gic. The advantage of this model lies in the fact that it considers anumber of factors that affect the input variables. However, man-agement of urban systems is a challenge because of the complexity

4248 G. Cirovic et al. / Expert Systems with Applications 41 (2014) 4245–4258

of those systems in circumstances which cannot always be exactlyforseen.

This paper presents a model for the routing of light deliveryvehicles by logistics operators. The model takes into account thefact that logistics operators have a limited number of EFV vehicles.When defining the routes, EFV and EUV are considered separately.An adaptive neural network is used for solving the problem of rout-ing in the model. The adaptive neural network was trained with asimulated annealing algorithm and was used to determine the per-formance of the branches of the network. The input parameters ofthe neural network were logistics operating costs and and the stateof the environmental parameters (exhaust emissions and noise) forthe given vehicle route. Each of the input parameters of the neuralnetwork was thoroughly examined. The input parameters werebroken down into elements that further describe the state of theenvironment and the logistics operating costs. The advantage ofthis model is reflected in the fact that it takes into account a great-er number of factors which affect the input variables. For example,when considering the input variable Exhaust emissions, the param-eters taken into account are those which describe oxides of sulfuremissions, nitrogen oxide emissions, carbon monoxide emissionsand particulate emissions. By means of mathematical transforma-tion the given parameters are brought together and they describethe input variable Exhaust emissions. In exactly the same way theremaining input variables of the neuro-fuzzy model (noise andlogistics operating costs) are broken down. After running the inputparameters through the neural network, the Peformance link (PL)is obtained at the output for every link in the network. Afterobtaining the PL values of the network the defined routes for lightdelivery vehicles are reached using a modified Clark–Wright algo-rithm. A detailed description of the individual phases of the modelis given in the following section of the paper. After describing thephases of the model the architecture of the developed adaptiveneural network is presented, the input parameters are definedand the process of training the network is shown. The final sectionof the paper shows how the model was tested on the routing oflight delivery vehicles in the old centre of Belgrade.

2. Neuro-fuzzy model for the routing of delivery vehicles bylogistics operators in urban areas

The problem presented in this paper is the problem of deter-mining a set of routes by means of which light delivery vehiclesin urban areas can provide their service in such a way that thelogistics operating costs and the state of the environmental param-eters are mimimal (Fig. 1b). Let there be n nodes within the

B

(a)Fig. 1. Nodes which require service and transport me

observed urban zone in which the delivery is made which areavailable to work. We mark demand with qi(i = 1,2, . . . ,n) in the ith node (Fig. 1a). The means of transport are stationed at point Bwhich is most commonly referred to as the base or depot. Allmeans of transport used for a particular task must begin and endtheir journey at point B. Let the capacity of the means of transpor-tation be greater than the demand at any node.

This problem is known as the standard vehicle routing problem.In order to make a more detailed mathematical formulation of therouting problem the following binary variables are introduced:

yik ¼1 if the kth means of transport services the ith node0 the opposite:

�

xik¼1 if the kth means of transport travels from node i to node j

0 the opposite:

�

We denote the performance link (PL) with cij in the network fromnode i to node j. Also, we use m to denote the number of vehiclesproviding the service, and Qk for the capacity of the kth means oftransport. The mathematical formulation of the problem of routingvehicles can be presented in the way it is described in the followingsection. It is necessary to minimize the criterion functionX

ijk

cijxijk ð1Þ

in accordance with the following constraints:

Xi

yik ¼1 i ¼ 2;3; . . . nm i ¼ 1

8><>: ð2Þ

Xik

qiyik 6 Qk; i ¼ 1;2; . . . ;n; k ¼ 1;2; . . . ;m ð3Þ

Xi

xijk ¼X

i

xijk ¼ yik; i ¼ 1;2; . . . ;n; k ¼ 1;2; . . . ;m ð4Þ

Xi;j2S

xijk 6 jSj � 1 8S # f2; . . . ;ng; k ¼ 1;2; . . . ;m ð5Þ

yik ¼ f0;1g i ¼ 1;2; . . . ;n; k ¼ 1;2; . . . ;m ð6Þ

xijk ¼ f0;1g i; j ¼ 1;2; . . . ;n; k ¼ 1;2; . . . ;m ð7Þ

The criterion function we want to minimize in this case is the sumof the total performance of the links in the network.

B

(b)ans and routes which complete the service task.

G. Cirovic et al. / Expert Systems with Applications 41 (2014) 4245–4258 4249

In this paper the given problem is presented using a neuro-fuz-zy model and a modified Clarke–Wright (CW) algorithm. The sched-uling of light delivery vehicles in urban areas on green routes is setout as a problem of optimization of logistics operating costs and airquality and noise level. The solution to the problem is proposedusing an Adaptive Neuro Fuzzy Inference System (ANFIS) whichtakes into account three criteria in the allocation of vehicles: x1 –logistics operating costs, x2 – exhaust emissions and x3 – noise le-vel. PLs are obtained as output from the ANFIS. PLs are obtainedthrough putting the input parameters (logistics operating costs, ex-haust emissions and noise) through an adaptive neural network.

Elements of Fig. 2: Input variables, determining the perfor-mance of the network links using a neuro-fuzzy model, Perfor-mance of the links in the network, Modified Clarke–Wrightalgorithm, Routing EFV vehicles, Routing EUV vehicles, Routes forEFV and EUV vehicles.

The scheduling of vehicles is carried out in two phases (Fig. 2).The First phase is the calculation of the input parameters of the AN-FIS and defining the PLs of the network.

In the second phase the PL values (cij) are assigned to thebranches in the network and the green routes are defined for EFVand EUV light delivery vehicles using a modified CW algorithm.EFV vehicle routing using the CW algorithm consists of the follow-ing steps:

Step 1. For each pair of nodes required to serve, calculate thesaving of the PL network (Cij) according to the expression (8)

Cij ¼ cBi þ cBj � cij ð8Þ

Fig. 2. Model for creating green routes for

Step 2. Carry out ranking of the savings performance and putthem in descending order (from the largest to the smallest).Make a list of savings which begin with the highest PL savings.Step 3. When considering the savings Cij include an appropriatebranch (i, j) in a partial route as long as it does not breach theexisting operational restrictions:(a) If none of the nodes i or j have been included in a partial

route.(b) If one of the nodes i or j has already been included in an

existing partial route and if that node is not an internalnode in the route.

(c) If both nodes i and j are included in two different partialroutes and neither of those nodes is internal to the routes.Then join the partial routes to become one route.

Step 4. When all the savings have been considered the algorithmis completed.

EUV vehicle routing using the CW algorithm consists of the fol-lowing steps:

Step 1. For every pair of nodes necessary for service calculate thesavings for the network performance links (C0ij) according toexpression (8).Step 2. Carry out ranking of the performance savings and putthem in ascending order (from the smallest to the largest).Make a list of the savings which begins with the lowest PLsavings.

After calculating the modified savings, steps 3 and 4 are re-peated as described in EFV vehicle routing.

light delivery vehicles in urban areas.

P

P

P

P

P

P

P

P

P

NORM

ALIZ

ATIO

N

1

4

2

3

5

6

7

8

9

x

y

∑

X1

X2

X3 P

P

P

P

P

P

15

16

17

18

19

20

М

VL

L

H

VH

М

VL

L

H

VH

М

VL

L

H

VH

y

1iO

2iO

3iO

4iO

5iO

Fig. 3. A five layered feedforward adaptine neural network.

Table 1Domain intervals for x1, x2, x3 and y.

Variable Domain

x1 [0,1]x2 [0,1]x3 [0,1]y [0,10]

4250 G. Cirovic et al. / Expert Systems with Applications 41 (2014) 4245–4258

The following section of the paper describes the process ofdefining the input parameters of the ANFIS and the architectureof the adaptive neuro-fuzzy network.

3. Configuration of the neuro-fuzzy network and defining theinput parameters: neuro-fuzzy modeling

An integral part of the ANFIS is a fuzzy logic system (FLS) of rea-soning. One of the basic problems facing the analyst when develop-ing an FLS is determining the set of linguistic rules anddetermining the parameters of the membership functions of the in-put/output pairs. The initial fuzzy system is mapped into a five lay-ered adaptive neural network with a restricted connectivitystructure that is shown in Fig. 3.

The proposed neural network is referred to as a five layered net-work because five layers perform operations. The adaptive networkshown in Fig. 3 is a feedforward layered network because the out-put of each unit propagates from the input side (left) to the outputside (right).

Based on analysis of the given literature references, three crite-ria were identified which affect the routing of delivery vehicles inurban areas, that is, which influence the definition of the PLs in thenetwork. The input variables of the ANFIS are: Logistics operatingcosts (LOC), Exhaust emissions (EE) and Noise (N). In addition tothe three input variables, the ANFIS has one output variable (y)The link performance. The intervals of the input and output vari-ables of the ANFIS are shown in Table 1.

The mathematical formulation of the input variables of the AN-FIS is given in the following way. The criterion Logistics operatingcosts (x1) is determined using the following expression (9)

fx1 ¼ Ckm � li þ Cn � tpb ð9Þ

where Ckm represents the vehicle operating costs per kilometer; lithe Eucledian length of the link; Ch – personnel costs of individualvehicles per hour of driving; tpb – time spent driving the vehiclein the link.

The criterion Exhaust emissions (x2) is obtained using the expres-sion (10)

fx2 ¼ x1 � SO2 þx2 � COþx3 �NOX þx4 � ðPM10 þ PM2:5Þ ð10Þ

where xi represents the weighting factor of the components forassessing the air quality – SO2, CO, NOX, PM10 and PM2.5. When con-sidering the parameters for assessing the air quality as representa-tive chemical compounds describing the state of the air, thecompounds SO2, CO, NOX, PM10 i PM2.5 were chosen because of theirharmful effects on the environment. Namely, the quantity of sulfuroxides emitted directly depends on their content in the fuel and thecombustion mode of the engine, and the harmful effects are re-flected in the acidification of the existing ecosystem. Emission ofnitrogen oxides was chosen because of their multiple harmful ef-fects on the ecosystem (acidification of the environment, destruc-tion of ozone in the upper layers of the atmosphere, etc.). Carbonmonoxide was chosen because of its strong cytotoxicity to livingbeings and also because it is one of the greatest air polluters. The

Table 2Limit values of the input parameters.

Input parameters Ckm (n.j./km) Cn (n.j./h) tpb (min) Lb (dB)

Limit values (50–100) (200–250) (3–10) (50–120)

G. Cirovic et al. / Expert Systems with Applications 41 (2014) 4245–4258 4251

exhaust emissions of internal combustion engines are one of thebiggest polluters of the atmosphere of this gas. Particulate matterwas chosen since it consists of a mixture of solid and liquid particlesof both organic and non-organic substances which, as a complexmixture, have a very negative effect on the human organism (byinhalation they are introduced and deposited in the respiratory sys-tem). The main components of particulate matter are sulfates, ni-trates, ammonia, sodium chloride, carbon, mineral dust and water(SEA, 2012; SEA, 2013). The concentration of particulate matter inthe ambient air is generally quantified today by measuring the con-centration of PM10 (particles with a diameter of less than 10 lm) orPM2.5 (diameter less than 2.5 lm).

The Noise (x3) criterion. A theoretical solution for the problem ofmodeling the level of traffic noise is very complicated due to thelarge number of different variables that affect the noise, and becauseof the lack of analytical equations that describe the correlation rela-tionships between the levels of noise and individual factors whichaffect it. The developed theoretical models include the characteris-tics of the noise source for calculating the emission of the noisesource and modeling the propogation from the place where theyare emitted to the place of noise immission, that is, the calculationpoint. Theoretical models are more precise, but their calculation istime-consuming and they are only used for the formation of engi-neering models. For this reason, mathematical models are devel-oped which are based on experimental results of measuring thenoise level and establishing correlations with the traffic parameters.

Many authors have offered a large number of mathematicalmodels (linear and nonlinear, statistical, based on fuzzy logic andneural networks) to describe traffic noise, with different levels ofaccuracy and which differ in terms of the factors they take into ac-count. All available models are based on establishing a functionalrelationship between the parameter of noise emissions and theparameters of traffic and roads. Some of the most widely used havebeen defined by Burgess (1977), (Eq. (11)), Josse (1972), (Eq. (12))and Fagotti & Poggi (1995), (Eq. (13)):

Leq ¼ 55:5þ 10:2 log Q þ 0:3p� 19:3 logðL=2Þ ð11Þ

Leq ¼ 38:8þ 15 log Q � 10 log L ð12Þ

Leq ¼ 10 logðNc þ Nm þ 8Nhv þ 88NbÞ þ 33:5 ð13Þ

where: p – is the percentage of heavy vehicles, L – is the road width(in meters), Q – is the total number of vehicles per hour, Nc – is thenumber of light (passenger) vehicles per hour, Nm – is the number ofmotorcycles per hour, Nhv – is the number of heavy (goods) vehiclesper hour, Nb – is the number of buses per hour. The total number ofvehicles in an hour (Q) in the above equations is expressed as anequivalent number of vehicles and is obtained under the assump-tion that one heavy vehicle is equivalent to 6 light vehicles, andone motorcycle to 3 light vehicles.

The majority of the mathematical models are obtained on thebasis of experimental data. This has the consequence that eachmodel in itself includes certain features of the place where the datawas collected and characteristics specific to that particular trafficflow. This does not ensure an accurate approximation of the trendof change for the equivalent noise level depending on a set numberof physical paramaters which define traffic and roads. Thereforeeach model should be very cautiously applied to the conditionsthat apply in other urban areas. For this reason, there is a needfor the development of a mathematical model which calculatestraffic noise whilst taking into account the structure of the vehiclesand the characteristics of the roads in specific urban areas.

Serbia does not have a developed model for calculating road traf-fic noise emissions, but instead uses software packages for noise cal-culation or other available models. Using an already existing modelfor modeling the level of traffic noise in the territory of Belgrade does

not have characteristic universality since each model contains cer-tain characteristics of the place where the measurements were ta-ken and specific features of the traffic flow of the urban areastudied. For the purpose of this research, the model for determiningequivalent noise levels of road transport was used, developed byPrascevic and Cvetkovic (2013). This model was developed withinthe framework of their research methodology for determiningequivalent noise levels for road traffic in the city of Niš.

fx3 ¼ 10logðNcþ3:7Nhv þ l:9NbÞþ38:2; 55 dBðAÞ< fx3 < 65 dBðAÞ ð14Þ

fx3 ¼10logðNcþ11:7Nhv þ3:1NbÞþ44:3; 65 dBðAÞ< Leq<120 dBðAÞ ð15Þ

where the number of passenger vehicles is (Nc), the number of hea-vy vehicles is (Nhv) and the number of buses is (Nb). The equivalentnoise level of road traffic at a distance of 7.5 m from the road wasdetermined on the basis of the number of passenger vehicles, thenumber of heavy vehicles and the number of buses in an hour.Based on the structure of traffic flow, the equivalent level of roadtraffic was calculated using Eq. (14). If the noise level obtainedwas less than or equal to 65 dB(A) then the noise level was recalcu-lated using Eq. (15).

The limit values of the input parameter variables fx1 , fx2 and fx3

which are characterized as the greatest and highest allowed valuesof the given parameters were defined on the basis of recommenda-tions by the Serbian Environmental protection agency, followingthe recommendations by Tao and Xinmiao (1998). Table 2 showsthe parameter values proposed for Serbia.

The parameter values shown depend on the economic situationand economic policies of each country where the proposed algo-rithm would be applied. On the basis of data from the SerbianAgency for environmental protection the defined limit values areSO2[0,125] lg/m3, CO[0,10] lg/m3 (maximum annual eight-hourvalue) and NOX[0, 85] lg/m3. The World Health Organization(WHO, 2012), in the Guidelines for the limit values of particulatematter, does not define the level of concentration (PM10 i PM2.5)as the lower threshold, but rather, it is the concentration belowwhich there is no impact on human health. The defined values areonly the recommended values which represent the concentrationsthat can be realistically achieved in order for the effects on humanhealth to be kept to a minimum. Table 3 presents an overview of thelimit values for the concentration of PM2.5, PM10, BS (soot) and TSP(total suspended particles) defined by the regulations of the Repub-lic of Serbia (Law on Ambient air protection., 2009; the Bylaw onmonitoring conditions and air quality requirements, 2010), by EUDirective (Directive on Ambient Air Quality, 2008) and therecommendations of the WHO (2012).

In order for the criteria functions (fx1 ; fx2 ifx3 ) to go through theANFIS it is necessary to carry out their normalization. Since fx1 be-longs to the group of cost criteria (lower values desirable) the nor-malization process is carried out according to the formula (16)

fxn ¼ 1� fxi� f min

x

f maxx

ð16Þ

The process of normalization for benefit criteria (higher valuesdesirable) fx2 i fx3 is carried out according to the formula (17).

fxn ¼fxi

f maxx

ð17Þ

Table 3Limitation values according to Serbian legislation, EU directives and WHO recommendations.

Period Republic of Serbia EU WHO

PM10 PM2.5 BS TSPa PM2.5 PM10 PM2.5 PM10

1 h (lg/m3) – – 150 – – – – –24 h (lg/m3) 50 – 50 120 – 50 25 501 year (lg/m3) 40 20 50 – 20 40 10 20

a In this paper the limit value of the total suspended particulate matter has a value of 120 lg/m3 (WHO, 2012).

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

0.2

0.4

0.6

0.8

1VL L M H VH

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

0.2

0.4

0.6

0.8

1

Exhaust emissions

Deg

ree

of m

embe

rshi

p

VL L M H VH

Deg

ree

of m

embe

rshi

p

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

0.2

0.4

0.6

0.8

1

Logistics operating costs

VH H M L VL

Deg

ree

of m

embe

rshi

p

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

0.2

0.4

0.6

0.8

1

Deg

ree

of m

embe

rshi

p

VH H M L VL

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

0.2

0.4

0.6

0.8

1

Noise

Deg

ree

of m

embe

rshi

p

VH H M L VL

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0

0.2

0.4

0.6

0.8

1

Noise

Deg

ree

of m

embe

rshi

p

VH H M L VL

(a) Membership functions of the input variables – before training

(b) Membership functions of the input variables – after training

Exhaust emissions Logistics operating costs

Fig. 4. Membership functions of the input variables – before training (first layer) and after training (second layer) of the ANFIS .

4252 G. Cirovic et al. / Expert Systems with Applications 41 (2014) 4245–4258

3.1. Pattern of connectivity

The neural network is designed to establish and compute afunction from input space to output space. In this paper, the net-work that has a fixed structure is configured based on the opera-tion of the fuzzy system (the Takagi–Sugeno model). The inputlayer consists of 3 units representing: Logistics operating costs(x1), Exhaust emissions (x2) and noise (x3). It simply transfers in-puts further via the interconnections to the hidden or first layer.All units in the input layer (x1, x2 and x3) are connected with the5 units in the first layer. The strengths of connections betweenthe units in the input layer and the units in the first layer are crispnumbers equal to 1.

The first layer consists of 3 + 5 units representing the number ofverbal descriptions quantified by fuzzy sets (‘‘very low’’, ‘‘low’’,‘‘medium’’, ‘‘high’’, ‘‘very high’’) for each input variable (Fig. 4).Every unit in the first layer is an adaptive unit with its output beingthe membership value of the premise part.

The number of units in the second layer equals the number offuzzy rules. Every unit in this layer is a fixed unit that calculatesthe minimum value of the two incoming inputs. The outputs fromthis layer fire strengths of rules. For example, the output from thefirst unit in the second layer is

w1 ¼minflLðx1Þ;lVHðx2Þ;lMðx3Þg

The third layer has five adaptive units representing the strength ofthe preference (‘‘very weak’’, ‘‘weak’’, ‘‘medium’’, ‘‘strong’’, ‘‘verystrong’’). Each unit in this layer calculates the intersection of a fuzzyset (consequent) with the maximum firing strength of incomingrules. For example, the fourth unit calculates the intersection of afuzzy set ‘‘strong preference’’ with the maximum firing strengths ofrules R5, R6, R7: lV4(y) = min{wi,lSP(y)}, where wi = max{w5,w6,w7}.The single unit in the fourth layer is a fixed unit that computes theoverall output of the ANFIS:

lMðyÞ ¼maxflV1ðyÞ;lV2ðyÞ;lV3ðyÞ;lV4ðyÞ;lV5ðyÞg ð18Þ

The obtained output is then defuzzified in the single unit in thefifth layer. Selection of the final crisp value can be made in variousways. In this paper the action which is closest to the center of grav-ity has been computed (Center-of-Gravity method). The output va-lue is a real number that lies in the interval [0,1]:

O ¼ Overalloutput ¼X

i

fwi fi ¼P

iwifiPiwi

ð19Þ

3.2. Supervised learning (a simulated annealing algorithm)

The aim of learning is to set the membership functions of the in-put/output variables to some adequate functions. The neural

G. Cirovic et al. / Expert Systems with Applications 41 (2014) 4245–4258 4253

network performances are measured as the deviation between thetargeted output and the model output across all numerical exam-ples. This discrepancy or the error measure is considered as theobjective function, and heuristic simulated annealing is used tominimize it. Since the application of simulated annealing requiresa large number of experiments, the training process is very long.However, the tuned FIS yields results superior to those obtainedby the initial fuzzy controllers and can be used in real time.

In this paper, the objective function that has to be minimized iscalculated as the sum of differences between the model output(maximum of the networks outputs) and the targeted output overall training pairs. Heuristic simulated annealing is used to mini-mize the objective function. A statistical training method such assimulated annealing requires the definition of an energy function(objective function) depending upon the parameters of the neuralnetwork. Whenever a new set of membership functions is gener-ated randomly, the resulting energy is determined. If the obtainedenergy is improved, then a new set of membership functions ismemorized, otherwise the acceptance or the rejection of thechange is decided according to a given probability distribution.The possibility that a change that worsens (increases) the energyis retained implies that the algorithm would hardly be trapped inlocal energy minima.

In this research, the fact that the algorithm would hardly betrapped in local minima, i.e., it will converge to a global minimum,was the main reason to choose this method as a learning rule. Thedisadvantage of using the simulated annealing algorithm as alearning rule for the neural network is its very long training period.

The simulated annealing algorithm based on the approaches ofKirkpatrick, Gellat, and Vecchi (1983) and Golden and Skicsim(1985) consists of the following steps:

Step 1. Develop a proper annealing schedule {t1, t2, . . . , tk} con-sisting of a sequence of temperatures (control parameters), t1 -> t2 > � � � > tk, and the amount of time required to reach theequilibrium at each temperature. Set i = 1.Step 2. Generate a set of initial membership functions. Obtainthe model output for all the input vectors in a training set (max-imum of the network outputs) and calculate the objective func-tion value (FO-old). The objective function minimizes the sum oferrors between the targeted outputs and the model outputs.Step 3. Generate a new set of membership functions by a smallperturbation. Obtain the model output (maximum of the net-work outputs) with the same input vectors (for the completetraining set) and calculate the new objective function value(FO-new). Evaluate the change in the objective function (d = FO--new � FO-old). If d < 0, go to Step 5, otherwise go to Step 4.Step 4. (d P 0) Compare a random variable (r) drawn from a uni-form distribution on the [0,1] interval, with the probability ofaccepting the new set of membership functions P(d) = exp(�d/ti). If r < P(d), go to Step 5, otherwise keep the old set of member-ship functions, and go to Step 3.Step 5. (d < 0 or r < P(d)) Memorize the new set of membershipfunctions and the new objective function value.

Table 4Parameters of the membership functions before and after training the ANFIS model.

Input value MF 1 (trimf) MF 2 (trimf)

Before trainingLOC [�0.026,0.102,0.239] [0.126,0.296,0.465]EE [�0.018,0.123,0.284] [0.131,0.278,0.426,8]N [�0.047,0.100,0.254,] [0.123,0.292,0.460]

After trainingLOC [�0.418,�0.018,0.382] [�0.124,0.252,0.628]EE [�0.394,�1.388e�017,0.394] [�0.131,0.238,0.607]N [�0.244,0.019,0.283] [�0.127,0.195,0.518]

Step 6. If the thermal equilibrium has been reached at the tem-perature ti, set i = i + 1. Steady state or equilibrium is reachedwhen we observe that an improvement of the objective func-tion is highly unlikely. An epoch is an interval between check-ing if the equilibrium is reached. The epoch implies kexchanges of all membership functions, where k is a pre-defined number. The best solution, i.e., the lowest sum of errorsbetween the model outputs and targeted outputs, obtainedthrough k exchanges of the membership functions, representsthe epoch. Consider the case where k epochs have already beengenerated. After the next epoch, equilibrium is reached if:

jFO�Vkþ1� FO�Vp j=FO�Vkþ1

< e; k ¼ 1; . . . ;Mep � 1 ð20Þ

where FO�Vkþ1is the objective function value that represents the

k + 1st epoch, FO�Vp is the lowest objective function value of all pre-vious epochs’ solutions and e, a pre-defined constant.

The maximum number of generated epochs at one temperatureif the thermal equilibrium is not reached in the meantime, Mep, isset in advance.

If i > K, the algorithm is completed. The solutions obtained by asimulated annealing algorithm do not depend on the initial solu-tion and usually approximate the optimal solution. However, theannealing schedule, i.e., the way the temperature gradually de-creases and the initial temperature both influence the performanceof the algorithm. Initially, a temperature is given a high value;then, it is slowly reduced until a small value, for which no deteri-orations are accepted any more, is reached. Thus, the convergenceof the obtained values for the membership functions is inheritedfrom the convergence of the simulated annealing algorithm.

While training the ANFIS there was a change in the parametersof the membership function of the input variables (Fig. 3b). Aftercompleting the final phase of training the ANFIS, the final parame-ters of the membership function were obtained (Table 4).

The application of simulated annealing requires a large numberof experiments. The initial temperature and the number of temper-atures are varied during numerous experiments. The best result isobtained when an array of 55 temperatures (ti+1 = 0.85ti) and aninitial temperature t1 = 55 are used.

The maximum number of generated epochs at one temperatureMep = 40. In other words, 40 epochs are generated if the thermalequilibrium is not reached in the meantime. The epoch implied25 exchanges of all middle values of the membership functions.The value of e is used to check if the equilibrium reached is 0.07.

Training the ANFIS was carried out through five phases, each ofwhich had a total of 11 temperatures (total t55). The first trainingphase was completed after the first 11 temperatures, [0, t11]. Aftercompleting the first phase, the error obtained at the output had amean value of 1.371 (Fig. 5a).

In the following phase, after the next 11 temperatures ([t12–t22]), the error obtained at the output had a mean value of 0.984(Fig. 5b), which is a reduction of 28.22% in relation to the previousvalue. The third and fourth phases lasted for a total of 22 temper-atures ([t23–t44]). After the fourth phase of training, error at the

MF 3 (trimf) MF 4 (trimf) MF 5 (trimf)

[0.325,0.487,0.637] [0.529,0.687,0.847] [0.706,0.846,1.025][0.286,0.459,0.632] [0.481,0.653,0.840] [0.70, 0.846,1.011][0.304,0.502,0.701,] [0.559,0.722,0.909] [0.769,0.935,1.06]

[0.121,0.484,0.846] [0.304,0.716,1.129] [0.585,1.033,1.481][0.084,0.487,0.889] [0.343,0.75,1.157] [0.567,1 1.433][0.054,0.433,0.812] [0.318,0.742,1.166] [0.501,1.034 1.566]

0 50 100 150 2003

4

5

6

7

8

9

10

0 50 100 150 2004

5

6

7

8

9

10

0 50 100 150 2003

4

5

6

7

8

9

10

0 50 100 150 2004

5

6

7

8

9

10

a) Error = 1.371 ftrainingfANFIS ftraining

fANFIS

ftrainingfANFIS ftraining

fANFIS

b) Error = 0.984

d) Error = 0.319c) Error = 0.709

Fig. 5. Training data (I, II, III and IV phase) – ANFIS output .

0 50 100 150 2004

5

6

7

8

9

10

ftrainingfANFIS

Error = 0.146

Fig. 6. Training data (phase V) – ANFIS output.

4254 G. Cirovic et al. / Expert Systems with Applications 41 (2014) 4245–4258

output amounted to 0.319 (Fig. 5d), which compared to the previ-ous phase is a reduction of 55%. After the fifth and final phase, com-pleted after the last 11 temperatures ([t45–t55]), the error value atthe output of the model was 0.146 (Fig. 6). Upon completion of thefifth phase it was concluded that the error obtained at the outputwas acceptable. It can also be concluded that the network istrained and able to generalize new input data.

After training it can be observed that the system is sensitive andthat the output is gradual. The inert and over-sensitive parts of thesystem are removed, which would be the case with a fuzzy model(Fig. 7a). Presented in Fig. 7b is a set of output values of the ANFISafter training, that is, the scenario which describes the reaction ofthe system for individual input values.

The five-layered adaptive network was tested on 36 parameterswhich describe the network nodes, constructed to simulate the

prevailing conditions in the centre of Belgrade. The criteria valueswhich describe the given node were periodically put through theANFIS, in this way obtaining the PLs of the network nodes.

4. Testing the model

The consequences of intensive climate change in recent decadesare reflected in almost all parts of the world. The growing numberof catastrophes as a result of weather conditions is a cause for con-cern, but there is a willingness to try and reduce or prevent furtherclimate change. For this purpose it is necessary to reduce the emis-sion of gases which cause the greenhouse effect, and whose con-centration in the atmosphere has sharply increased, above all asa consequence of using fossil fuels in traffic and in industry. Inthe case of the transport sector, one of the main measures forachieving this goal is a redirection in the flow of goods towardsmore ecologically acceptable branches of transport and the use ofEFV vehicles.

Traffic congestion, which is very pronounced in the so-calledpeak traffic periods, is extremely harmful to the environmentand causes many negative consequences, such as: additional fuelconsumption, extra pollution, arriving late to work or school, sup-ply problems, a reduction in the productivity and effect of trans-port vehicles and others. In addition, the majority of harmfulgases (around 1/3) are a result of traffic, namely 65% of carbonmonoxide, 45% of hydrocarbons and 49% of nitrogen oxides (EEA,2013a, 2013b).

In order to inform the public about air quality, in Serbia theEnvironmental Protection Agency was established, which providesthe results of automatic air quality monitoring in real time. Formanaging the air quality it is important to have the data on currentlevels of air pollutants obtained by these measurements, as well as

0 0.2 0.4 0.6 0.8 1

00.5

1

4

5

6

7

8

Exhaust emissions Logistics operating costs

perf

eren

cija

0

0.5

1

00.2

0.40.6

0.81

4

5

6

7

8

Exhaust emissions Noise 00.2

0.40.6

0.81

0

0.5

1

5.5

6

6.5

7

7.5

8

NoiseLogistics operating costs

perf

eren

cija

00.2

0.40.6

0.81

0

0.5

1

5

6

7

8

9

Exhaust emissions Logistics operating costs

perf

eren

cija

00.2

0.40.6

0.81

0

0.5

1

5

6

7

8

9

Exhaust emissions Noise

perf

eren

cija

0 0.2 0.4 0.6 0.8 1

00.5

1

6

6.5

7

7.5

8

8.5

9

9.5

NoiseLogistics operating costs

perf

eren

cija

a) Sensitivity of the ANFIS - before training the model

b) Sensitivity of the ANFIS - after training the model

Fig. 7. Sensitivity of the ANFIS.

Table 5The emission of pollutants by motor vehicles in Belgrade.

Type of pollutant(after covering1000 km)

Passengercar

Means oftransport(capacity < 3.5t)

Means oftransport(capacity > 3.5t)

Bus

Averageconsumption(dm3)

104 43 140 549

Averageconsumption(kg)

73 30 98 386

Hydrocarbons(kg)

1.67 2.82 1.19 2.75

Benzene (kg) 0.092 0.025 0.054 0.245Carbon dioxide

(kg)218 937 299 1203

Carbon monoxide(kg)

9.86 4.26 7.04 9.84

Nitrogen oxides(kg)

0.864 11.5 1.41 19.0

Suspendedparticles (kg)

0.030 0.868 0.207 1.07

Abrasion rubber(kg)

0.064 0.768 0.112 0.768

Elemental carbon(kg)

0.028 6.41 0.141 0.694

Sulfur dioxide(kg)

0.047 0.960 0.210 1.23

G. Cirovic et al. / Expert Systems with Applications 41 (2014) 4245–4258 4255

data on the emission of harmful substances in the air which causethe pollution. For this purpose, 40 automatic measuring stationswere placed in Belgrade in order to monitor the quality of theair. Verified values and assessment of air quality according to zonesin Serbia are presented in annual SEA reports on the state of the airquality in the Republic of Serbia. The SEA report (2012) states thatthe greatest emissions of greenhouse gases are in Belgrade, NoviSad, Pancevo and Smederevo. The report also states that in Bel-grade, over 50% of the greenhouse gases are a result of exhaustemissions. Table 5 shows the emission of pollutants by motor vehi-cles in Belgrade during city driving (SEA, 2012). The data is basedon an individual motor vehicle after 1000 km of city driving.

Given the data presented in Table 5 and the SEA (2012) report,the local authorities in Belgrade have begun to subsidize the pur-chase of EFV in order to reduce the environmental pollution inthe city. The plan for the next four years is for logistics operatorsto have around 50% EFV in their fleets. However, in addition to pur-chasing EFV, logistics operators will face the problem of how toallocate EFV vehicles. For solving the problem of routing greenvehicles in Belgrade, the model developed in this paper can be suc-cessfully used.

The next section presents a neuro-fuzzy model implemented ona test network which simulates conditions in the centre of Bel-grade. Fig. 8 shows the test network, which consists of 9 nodes.The capacity of the vehicles which carry out the service isV = 130. The requirements of the nodes are shown beside the nodesas indicated in Fig. 8.

Input paramaters of the adaptive neural network (Table 6) wereobtained from the limit values for the period 2010–2013 (SEA,2013).

After defining the initial parameters of the network and calcu-lating the PLs using the ANFIS model, the algorithm for routingEFV is implemented.

The first phase is calculating the input parameters of the ANFISand defining the PLs for the network.

In the second phase PL values (cij) are assigned to the branches ofthe network and green routes for EFV and EUV light delivery

vehicles are defined using a modified CW algorithm. Defining aroute for EFV and EUV using the CW algorithm consists of the fol-lowing steps:

Step 1. Calculation of the input parameters for the ANFIS anddefinition of the network PLs using the neuro-fuzzy model(Table 6). When defining EFV and EUV routes for each pair ofnodes calculate the PL network savings (Cij i C0ij) according toexpression (8).Step 2. Carry out ranking of the PL savings and and put them indescending order (from the highest to the lowest) for the EFV

1

3

2

5

4

6

8

9

7

(40)(70) (30)

(60)

(20)

(30)(30)

(20)

Fig. 8. Network architecture for routing green vehicles.

Table 6Characteristics of the links in the tested network.

No. Link SO2 CO NOx PM10 PM2.5 FEIG li tpb fto fB PL

1. R1–2 55.11 6.12 68.12 56.2 78.5 0.78 4.0 7 0.36 0.53 8.452. R1–3 110.22 9 39.74 65.3 65.4 0.7 3.0 10.2 0.53 0.77 9.373. R1–4 20.04 9.16 68.12 41.2 54.7 0.89 5.8 7.6 0.39 0.57 9.044. R1–5 90.18 7.34 51.09 69.2 98.9 0.43 3.2 4.6 0.24 0.35 5.825. R1–6 75.15 7.06 79.47 44.9 114.2 0.57 5.2 8.5 0.44 0.64 8.086. R1–7 95.19 6.62 56.77 42.1 58.7 0.92 2.8 5.9 0.3 0.44 8.477. R1–8 85.17 6.1 45.41 37.9 58.3 0.82 6.7 3.9 0.2 0.29 7.598. R1–9 120.24 9.76 56.77 61.6 119.8 0.78 5.5 9.9 0.51 0.74 9.519. R2–3 25.05 7.52 39.74 112.5 120 0.51 4.3 10.9 0.57 0.82 8.6510. R2–4 55.11 9.48 39.74 79.6 112.9 0.79 8.1 7.4 0.38 0.56 8.6211. R2–5 105.21 7.38 39.74 103.6 74.7 0.47 2.9 6.4 0.33 0.48 6.5612. R2–6 90.18 7.84 51.09 35.7 60.3 0.57 8.7 5.2 0.27 0.39 6.7113. R2–7 100.2 6.12 34.06 76 56.9 0.48 6.3 3.4 0.17 0.26 5.8414. R2–8 90.18 7.46 68.12 61.3 61.8 0.86 10.6 7.2 0.37 0.54 8.8115. R2–9 115.23 8.22 45.41 82.5 114.3 0.62 9.4 5.7 0.29 0.43 7.1316. R3–4 70.14 7.54 51.09 57.7 99.4 0.68 3.7 6.2 0.32 0.47 7.6817. R3–5 90.18 9.9 79.47 69.9 82.5 0.87 5.3 3.5 0.18 0.26 7.7318. R3–6 50.1 9.0 34.06 86.7 100.7 0.55 4.6 6.1 0.31 0.46 6.8919. R3–7 125.25 7.96 51.09 107.8 67.2 0.67 5.7 5.5 0.28 0.41 7.3420. R3–8 120.24 9.04 39.74 72 45.0 0.90 7.0 10 0.52 0.75 9.7821. R3–9 55.11 7.56 51.09 89.3 69.0 0.50 7.0 3.7 0.19 0.28 5.9922. R4–5 125.25 6.2 51.09 93.1 106.2 0.48 8.8 4.5 0.23 0.34 6.0323. R4–6 70.14 9.54 34.06 65.9 102.2 0.67 2.7 5.4 0.28 0.41 7.3124. R4–7 50.1 9.36 79.47 87.7 81.6 0.83 7.7 4.6 0.24 0.35 7.8025. R4–8 110.22 9.12 79.47 51.2 92.1 0.63 5.7 3.8 0.19 0.29 6.6526. R4–9 55.11 7.22 73.8 58.4 57.0 0.61 6.7 8.6 0.44 0.65 8.3127. R5–6 30.06 9.9 45.41 116.2 52.5 0.46 8.4 7.0 0.36 0.53 6.7628. R5–7 40.08 6.64 79.47 107 69.3 0.82 4.0 4.0 0.2 0.3 7.6129. R5–8 85.17 8.44 45.41 101.3 108.4 0.96 9.6 5.5 0.28 0.41 8.4330. R5–9 80.16 6.02 51.09 72.2 82.3 0.86 8.0 8.2 0.42 0.62 9.1531. R6–7 45.09 9.3 39.74 113.1 93.6 0.97 6.2 8.1 0.42 0.61 9.4432. R6–8 15.03 7.5 51.09 70.2 40.6 0.59 3.0 6.4 0.33 0.48 7.2433. R6–9 45.09 9.62 34.06 92.9 112.3 0.47 4.3 6.4 0.33 0.48 6.5634. R7–8 100.2 6.78 73.8 96.8 83.1 0.92 6.2 4.6 0.24 0.35 8.1235. R7–9 110.22 7.14 68.12 107.4 84.1 0.94 4.0 7.5 0.39 0.56 9.1636. R8–9 120.24 7.02 39.74 51.6 82.0 0.73 2.5 5.0 0.26 0.38 7.48

4256 G. Cirovic et al. / Expert Systems with Applications 41 (2014) 4245–4258

vehicles. Carry out ranking of the PL savings for EUV vehiclesand arrange them in ascending order (from the lowest to thehighest). Make a list of savings that begins with the lowest sav-ings performance (for EUV vehicles) and the highest savingsperformance (for EFV vehicles), Table 7.Step 3. Determine the route between all pairs of nodes. Carryout a review of savings (Ciji C0ij) and construct routes for EFVand EUV vehicles whilst respecting the limitations defined ear-lier in this paper. After analyzing the savings from Table 7,

routes are constructed for EFV vehicles (Fig. 9a) and for EUVvehicles (Fig. 9b).

When defining routes it was assumed that the number of EFV islimited and that a maximum of one can be assigned to one of theroutes. Under these assumptions step 3 of the recommended algo-rithm was used and vehicles were assigned to the given routes.

5. Discussion and conclusions

In line with world trends, this paper contains a model devel-oped to optimize the implementation of EFV vehicles on the exist-ing road network taking into account air pollution, noise level andoperating and user costs. The criteria for modeling the user andoperator costs are in accordance with the literature mentionedhere and logical indicators of the given quantities, such as unit la-bor costs. For the criterion of environmental quality, substances aretaken into account which are byproducts of vehicle usage, andwhich are detrimental to the health of humans and animals, suchas noise level, which affects the general state of people and theirquality of life.

The parameters for calculating the input variables of the neuro-fuzzy system came from real data obtained from automatic mea-suring stations situated in Belgrade, and the model was tested ona network which simulates conditions in the centre of Belgrade.This testing could have been carried out using data from any city,which indicates it to be a widely applicable model. The model re-mains open for modification, as well as for further improvementand adaption to the requirements of the user, since it is possible,instead of the suggested method of obtaining input data, to use

Table 7Sorting the savings of the branches for EFV and EUV vehicles.

EFV vehicles EUV vehicles

No. Link Cij No. Link C0ij

1. R3–9 12.89 1. R5–8 4.982. R2–7 11.08 2. R5–9 6.183. R6–9 11.03 3. R5–7 6.684. R2–9 10.83 4. R6–7 7.115. R3–4 10.73 5. R5–6 7.146. R3–6 10.56 6. R3–8 7.187. R3–7 10.50 7. R2–8 7.238. R4–9 10.24 8. R3–5 7.469. R4–8 9.98 9. R2–5 7.7110. R2–6 9.82 10. R7–8 7.9411. R4–6 9.81 11. R6–8 8.4312. R4–7 9.71 12. R7–9 8.8213. R8–9 9.62 13. R4–5 8.8314. R2–3 9.17 14. R2–4 8.8715. R2–4 8.87 15. R2–3 9.1716. R4–5 8.83 16. R8–9 9.6217. R7–9 8.82 17. R4–7 9.7118. R6–8 8.43 18. R4–6 9.8119. R7–8 7.94 19. R2–6 9.8220. R2–5 7.71 20. R4–8 9.9821. R3–5 7.46 21. R4–9 10.2422. R2–8 7.23 22. R3–7 10.523. R3–8 7.18 23. R3–6 10.5624. R5–6 7.14 24. R3–4 10.7325. R6–7 7.11 25. R2–9 10.8326. R5–7 6.68 26. R6–9 11.0327. R5–9 6.18 27. R2–7 11.0828. R5–8 4.98 28. R3–9 12.89

1

5

8

6

43

9

7

2

1

5

8

6

43

9

7

2(a) Routes for EFV vehicles (b) Rute EUV vozila

Fig. 9. Routes for EFV and EUV vehicles.

G. Cirovic et al. / Expert Systems with Applications 41 (2014) 4245–4258 4257

other more advanced methods, which is particularly importantwhen considering the input criterion of Noise. Thus, instead ofthe suggested model of Prascevic and Cvetkovic (2013) used in thisstudy, it is possible to use any other which the user considers to beof good quality, that is, which better reflects the situation in ques-tion. Also, the structure of this model allows for the possibility ofthere being incomplete data. In other words, if there are no specificdata on the input parameters, the model still functions on the sameprinciple and gives valid results. This is very important when con-sidering the input parameters for the criterion of Emissions, whereusers will often be in a situation where they have incomplete datadue to a greater number of parameters, which cannot be easilymeasured. The application of the adaptive neural network makesit possible to continuously incorporate new theoretical and empir-ical knowledge that can be reached by using this and similar mod-els in practice.

In addition it should be noted that this is a general model for therouting of light commercial vehicles by logistics operators, which issuitable for use in cities which face the problem of allocating their

‘‘green capacity’’ in the town network. One of the advantages ofthis model is that it takes into account the uncertainties whicharise when predicting the operating costs, user costs and the envi-ronmental parameters in the city’s road network. In addition, theproposed model enables the planning of vehicle routes to maxi-mize the positive impact on the environment, which is reflectedin the reduction of harmful gas emissions and an increase in theair quality in areas with the highest concentration of population.The model for routing light commercial vehicles also takes into ac-count the fact that logistics distributors have a limited number ofEFV vehicles. Therefore, the model has specially construrctedroutes for EFV and EUV. Likewise, the algorithm supports any num-ber of available EFV and EUV, which is consistent with the size ofthe network and the transport requirements.

As far as the author is aware, there has been no research so faras to whether the position of fuel filling stations for vehicles using‘‘green’’ fuels would result in any changes in the proposed routes. Itis becoming more common to hear the idea of electric vehiclesbecoming part of the fleets of logistics operators. Future researchand improvement of the model should therefore take into accountthe problem of location for filling batteries on light delivery vehi-cles, and their optimal location in relation to the vehicle route.As a suitable technique for solving such a problem, in addition toneuro-fuzzy modeling, a mass serving system can also be used.

This model extends the theoretical framework of knowledge inthe field of green vehicles. The existing problem is consideredusing new methodology, which creates a basis for further theoret-ical, and also practical advancement. The presented model alsohighlights new criteria (exhaust emissions, noise and operatingcosts) which have not been considered in models up to this point,and which are relevant to this issue. By introducing and describingnew criteria in the model, the need for their consideration in fur-ther analysis of this and similar problems is pointed out.

This model has been developed to minimize air pollution, noiselevel and logistics operating costs. Precisely this makes it compli-ant with world trends, since traffic in general, but particularlygoods transport, has a very complex effect on the environment,resulting in a range of negative consequences, which can be seenin air pollution, water pollution, noise, energy consumption, re-duced safety, vibration and others. On the other hand, biofuelssuch as biogas, biomethane and natural gas are increasingly beingused and are replacing traditional fossil fuels. Although the bene-fits of using EFV are well-known, their introduction is gradual,and their optimal allocation on routes is very important, whichgives the model presented here great practical significance. Thepractical value of this algorithm lies in the fact that the collectedexperience of a number of experts is incorporated into the model,thus avoiding a situation in which the routing of EFV is limited tothe knowledge of individuals who find themselves in a positionwhere they have to solve these problems alone.

4258 G. Cirovic et al. / Expert Systems with Applications 41 (2014) 4245–4258

Acknowledgements

The work reported in this paper is a part of the investigationwithin the research project TR 36017 supported by the Ministryfor Science and Technology, Republic of Serbia. This support isgratefully acknowledged.

References

Baaj, M. H., & Mahmassani, H. S. (1995). Hybrid route generation heuristic algorithmfor the design of transit networks. Transportation Research Part C, 3, 31–50.

Bapna, R., Thakur, L. S., & Nair, S. K. (2002). Infrastructure development forconversion to environmentally friendly fuel. European Journal of OperationalResearch, 1423, 480–496.

Bektas, T., & Laporte, G. (2011). The pollution routing problem. TransportationResearch Part B, 45, 1232–1250.

Beltran, B., Carrese, S., Cipriani, E., & Petrelli, M. (2009). Transit network design withallocation of green vehicles: A genetic algorithm approach. TransportationResearch Part C, 17, 475–483.

Burgess, M. A. (1977). Noise prediction for urban traffic conditions – related tomeasurements in the sydney metropolitan area. Appllied Acoustics, 10, 1–7.

Bylaw on monitoring conditions and air quality requirements (2010). OfficialGazette Republic of Serbia, No. 11/10.

Carrese, S., & Gori, S. (2002). An urban bus network design procedure. In M.Patriksson, M. Labbè (Eds.), Transportation planning: state of the art (pp. 177–196).

Ceder, A., & Israeli, Y. (1993). Design and evaluation of transit routes in urbannetworks, In Proceedings of the 3rd international conference on competition andownership in surface passenger transport, Ontario, Canada, 1993.

Cipriani, E., Fusco, G., & Petrelli, M. (2006). A multimodal transit network designprocedure for urban areas. Journal of Advances in Transportation Studies,Section A, 10, 5–20.

Demir, E., Bektas, T., & Laporte, G. (2011). A comparative analysis of several vehicleemission models for road freight transportation. Transportation Research Part D,16(5), 347–357.

Directive on Ambient Air Quality and Cleaner Air for Europe (2008). Directive 2008/50/EC.

European enviroment agency (EEA) (2013). The impact of international shipping onEuropean air quality and climate forcing. Copenhagen, European environmentagency.

European enviroment agency (EEA) (2013). The impact of international shipping onEuropean air quality and climate forcing, Copenhagen, European environmentagency.

Erdogan, S., & Miller-Hooks, E. (2012). A green vehicle routing problem.Transportation Research Part E, 48, 100–114.

Fagotti, C., & Poggi, A. (1995). Traffic noise abatment strategies. The analysis of realcase not really effective. In Proceedings of 18th 1nternational congress for noiseabatement (pp. 223–233), Bologna, Italy.

Fan, W., & Machemehl, R. B. (2006). Optimal transit route network design problemwith variable transit demand: Genetic algorithm approach. Journal ofTransportation Engineering, 132(1), 40–51.

Faulin, J., Juan, A., Lera, F., & Grasman, S. (2011). Solving the capacitated vehiclerouting problem with environmental criteria based on real estimations in roadtransportation: A case study. Procedia Social and Behavioral Sciences, 20,323–334.

Figliozzi, M. (2010). Vehicle routing problem for emissions minimization.Transportation Research Record, 2197, 1–7.

Frey, C., Zhang, K., & Rouphail, N. (2010). Vehicle-specific emissions modeling basedupon road measurements. Environmental Sciences and Technology, 44(9),3594–3600.

Golden, B., & Skicsim, C. (1985). Using simulated annealing to solve routing andlocation problems. Working paper MS/S 85-010, University of Maryland atCollege Park.

Jabali, O., Van Woensel, T., & de Kok, A. (2012). Analysis of travel times and CO2

emissions in time-dependent vehicle routing. Production and OperationsManagement, 21(6), 1060–1074.

Jaramillo, J. (2010). The green single vehicle routing problem. SE INFORMS AnnualMeeting, Myrtle Beach, SC.

Josse, R. (1972). Notions d’acoustique Paris. France. Ed. Eyrolles.Kara, I., Kara, B., & Yetis, M. (2007). Energy minimizing vehicle routing problem. In

A. Dress, Y. Xu, & B. Zhu (Eds.), Combinatorial optimization and applications(pp. 62–71). Berlin, Heidelberg: Springer.

Kirkpatrick, S., Gellat, L., & Vecchi, M. (1983). Optimization by simulated annealing.Science, 220, 671–680.

Kuo, Y. (2010). Using simulated annealing to minimize fuel consumption for thetime-dependent vehicle routing problem. Computer and Industrial Engineering,59, 157–165.

Kuo, Y., & Wang, C. C. (2011). Optimizing the vrp by minimizing fuel consumption.Management of Environmental Quality: An International Journal, 22(4), 440–450.

Law on Ambient air protection (2009). Official Gazette Republic of Serbia, No. 36/09.Lin, C., Choy, K. L., Ho, G. T. S., Chung, S. H., & Lam, H. Y. (2014). Survey of green

vehicle routing problem: Past and future trends. Expert Systems withApplications, 41, 3189–3203.

McKinnon, A. (2008). Logistical restructuring, freight traffic growth and theenvironment, London.

Murphy, P., & Poist, R. (2003). Green perspectives and practices: A comparativelogistics study. International Journal of Supply Chain Management, 8(2), 122–131.

Ngamchai, S., & Lovell, D. J. (2003). Optimal time transfer in bus transit routenetwork design using a genetic algorithm. Journal of Transportation Engineering,129(5), 510–521.

Ntziachristos, L., & Samaras, Z. (2000). Copert iii computer programme to calculateemission from road transport: Methodology and emission factors. Technicalreport No 49, European Environment Agency, Copenhagen, Denmark.

Peng, Y., & Wang, X. (2009). Research on a vehicle routing schedule to reduce fuelconsumption. In Proceedings of 2009 international conference on measuringtechnology and mechatronics automation (pp. 825–827), IEEE.

Prascevic, M., & Cvetkovic, D. (2013). Verification of NAISS model for road trafficnoise prediction in urban areas. Elektronika ir elektrotechnika, 19(6), 91–94.

Sbihi, A., & Eglese, R. (2007). The relationship between vehicle routing andscheduling and green logistics – a literature survey. Working Paper. TheDepartment of Management Science, Lancaster University.

Sbihi, A., & Eglese, R. (2010). Combinatorial optimization and green logistics. Annalsof Operations Research, 175(1), 159–175.

Schreyer, C., Schneider, C., Maibach, M., Rothengatter, W., Doll, C., & Schmedding, D.(2004). External costs of transport. Zurich/Karlsruhe: Update study.

Scott, C., Urquhart, N., & Hart, E. (2010). Influence of topology and payload on CO2

optimized vehicle routing. In Proceedings of the 2010 international conference onapplications of evolutionary computation (Vol. 2, pp. 141–150), Springer, Berlin,Heidelberg.

Serbian environment agency (SEA) (2012). Environment in Serbia. Belgrade, Serbia,Serbian environment agency.

Serbian enviroment agency (SEA) (2013). National network for monitoring airquality. Serbia, Serbian enviroment agency.

Shi, Y., & Mizumoto, M. (2000). A new approach of neuro-fuzzy learning algorithmfor tuning fuzzy rules. Fuzzy Sets and Systems, 112(1), 99–116.

Shing, J., & Jang, R. (1993). ANFIS: Adaptive network – based fuzzy inference system.IEEE Transactions on Systems, Man, and Cybernetics, 23(3), 665–685.

Shing, J., & Jang, R. (1995). Neuro – fuzzy modeling and control. The Proceeding of theIEEE, 83(3), 378–406.

Suzuki, Y. (2011). A new truck-routing approach for reducing fuel consumption andpollutants emission. Transportation Research Part D, 16, 73–77.

Tao, Y., & Xinmiao, Y. (1998). Fuzzy comprehensive assessment, fuzzy clusteringanalysis and its application for urban traffic environment quality evaluation.Transportation Research D, 3(1), 51–57.

Ubeda, S., Arcelus, F., & Faulin, J. (2010). Green logistics at Eroski. InternationalJournal of Production Economics, 131(1), 1–8.

WHO (2012). Air quality guidelines for particulate matter, ozone, nitrogen dioxideand sulfur dioxide, Summary of risk assessment.

Xiao, Y., Zhao, Q., Kaku, I., & Xu, Y. (2011). Development of a fuel consumptionoptimization model for the capacitated vehicle routing problem. Computers andOperations Research, 39(7), 1419–1431.

Yazan, D. M., Petruzzelli, A. M., & Albino, V. (2011). Analyzing the environmentalimpact of transportation in reengineered supply chains: A case study of aleather upholstery company. Transportation Research Part D, 16, 335–340.