Environmental Impact Assessment Review 49 (2014) 24–30

Contents lists available at ScienceDirect

Environmental Impact Assessment Review

j ourna l homepage: www.e lsev ie r .com/ locate /e ia r

Comparison of classical statistical methods and artificial neural networkin traffic noise prediction

Vladimir Nedic a,1, Danijela Despotovic b,2, Slobodan Cvetanovic c,3, Milan Despotovic d,⁎, Sasa Babic e

a Faculty of Philology and Arts, University of Kragujevac, Jovana Cvijića bb, 34000 Kragujevac, Serbiab Faculty of Economics, University of Kragujevac, Djure Pucara Starog 3, 34000 Kragujevac, Serbiac Faculty of Economics, University of Niš, Trg kralja Aleksandra Ujedinitelja, 18000 Niš, Serbiad Faculty of Engineering, University of Kragujevac, Sestre Janjic 6, 34000 Kragujevac, Serbiae College of Applied Mechanical Engineering, Trstenik, Serbia

⁎ Corresponding author. Tel.: + 381 69 844 96 79.E-mail addresses: vnedic@kg.ac.rs (V. Nedic), ddespot

slobodan.cvetanovic@eknfak.ni.ac.rs (S. Cvetanovic), mde(M. Despotovic), babicsf@yahoo.com (S. Babic).

1 Tel.: +381 63 81 35 373; fax: +381 34 304 275.2 Tel.: +381 63 40 80 23; fax: +381 34 303 516.3 Tel.: +381 63 42 43 11; fax: +381 18 293 932.

http://dx.doi.org/10.1016/j.eiar.2014.06.0040195-9255/© 2014 Elsevier Inc. All rights reserved.

a b s t r a c t

a r t i c l e i n f o

Article history:Received 11 April 2014Received in revised form 16 June 2014Accepted 23 June 2014Available online xxxx

Keywords:Artificial neural networkSoftwareOptimizationTraffic noise

Traffic is the main source of noise in urban environments and significantly affects human mental and physicalhealth and labor productivity. Therefore it is very important to model the noise produced by various vehicles.Techniques for traffic noise prediction are mainly based on regression analysis, which generally is not goodenough to describe the trends of noise. In this paper the application of artificial neural networks (ANNs) forthe prediction of traffic noise is presented. As input variables of the neural network, the proposed structure ofthe traffic flow and the average speed of the traffic flow are chosen. The output variable of the network is theequivalent noise level in the given time period Leq. Based on these parameters, the network is modeled, trainedand tested through a comparative analysis of the calculated values and measured levels of traffic noise usingthe originally developed user friendly software package. It is shown that the artificial neural networks can be auseful tool for the prediction of noise with sufficient accuracy. In addition, the measured values were also usedto calculate equivalent noise level by means of classical methods, and comparative analysis is given. The resultsclearly show that ANN approach is superior in traffic noise level prediction to any other statistical method.

© 2014 Elsevier Inc. All rights reserved.

Introduction

In recent years, due to the constant increase of population and thenumber of circulating vehicles in urban areas, pollution reached analarming level. Apart from air pollution, a very important factor regard-ing environmental pollution in urban areas is noise. Among differentsources of noise that are present in an urban area, traffic noise is by farthe most annoying noise source (Calixto et al., 2003). The influence oftraffic noise on human health has been studied numerously in recentyears (Babisch et al., 2013; Brink, 2011; Caciari et al., 2013; Fyhri andKlboe, 2009; Pirrera et al., 2010), the results of which confirmed thatthis kind of annoyance significantly affects both mental and physicalhealth. Therefore, traffic noise is to be considered not only as a causeof nuisance, but also as a concern for public health and environmentalquality (Kassomenos et al., 2014). To successfully implement the mostefficient noise action plans for preventing and reducing exposure toharmful levels of noise in a sustainable and resource efficient way, it is

ovic@kg.ac.rs (D. Despotovic),spotovic@kg.ac.rs

first necessary to obtain information about the noise levels to whichpeople are exposed (Suarez and Barros, 2014; Kassomenos et al.,2014). Thus, in order to control noise sound level in urban areas, it isvery important to develop methods for prediction of the traffic noise.The first traffic noise prediction (TNP) models date back to early1950s. Since then a large number of methods and models for trafficnoise prediction have been developed. The critical reviews of the mostused ones are given in Steele (2001) and Quartieri et al. (2009) aswell as in Garg and Maji (2014). Most of the TNP models that are pre-sented in literature are based on linear regression analysis. The mainlimit of those models, as concluded in Quartieri et al. (2009) andClaudio Guarnaccia et al. (2011), is “that they don't take into accountthe intrinsic random nature of traffic flow, in the sense that they don'ttake care of how vehicles really run, considering only how many theyare”. More advanced models involve artificial neural networks (ANN)(Cammarata et al., 1995; Givargis and Karimi, 2010) and genetic algo-rithms (Güdogdu et al., 2005; Rahmani et al., 2011). ANN model thatwas used in Cammarata et al. (1995) has 3 inputs: equivalent numberof vehicles, which was obtained by adding to the number of cars thenumber of motorcycles multiplied by 3 and the number of trucksmultiplied by 6, the average height of the buildings on the sides of theroad, and the width of the road. In order to increase the number ofinputs the authors decomposed equivalent number of vehicles intothe number of cars, the number of motorcycles, and the number of

Fig. 1. Schematic representation of neural network.

25V. Nedic et al. / Environmental Impact Assessment Review 49 (2014) 24–30

trucks, and got the ANN model with 5 inputs. The principle of neuralarchitecture consists of two phases, first of which is the filtering ofacoustic measurements affected by error by means of learning vectorquantization (LVQ) network. The model is tested on measured dataset, and also compared with three classical models (Burgess, Josse, andCSTB). It was found that agreement between predictions and measure-ments was much better for the neural network approach than for theclassical ones. However, the main drawback of that approach usingLVQ network is, as authors point out, its dependency on the data usedin the training phase. That creates difficulties that network trained forone town layout (road width and building height), be applied for atownwith a totally different layout. In terms of the parameters involvedin the CoRTN (calculation of road traffic noise) model (Quartieri et al.,2009), which was initially developed in 1975 by the Transport andRoad Research Laboratory and the Department of Transport of theUnited Kingdom, the ANN model that was used in Givargis and Karimi(2010) has 5 input variables: the total hourly trafficflow, the percentageof heavy vehicles, the hourly mean traffic speed, the gradient of theroad, and the angle of view. The authors tested the developed modelon the data collected on Tehran's roads, and found no significant differ-ences between the outputs of the developed ANN and the calibratedCoRTN model.

In this paper an application of artificial neural networks for theprediction of traffic noise is presented. The developed ANN model has5 input variables: the number of light motor vehicles, the number ofmedium trucks, the number of heavy trucks, the number of buses andthe average traffic flow speed. The network is modeled, trained andtested on datameasured on Serbian road using the originally developeduser friendly software package. Furthermore, the comparison betweenthe outputs of the developed network and the outputs of some classicalmethods is given. As it will be shown, the developed ANN model hasmuch better capabilities to predict traffic noise level than any otherclassical method.

Problem formulation

The most suitable parameter for depicting traffic noise emission isequivalent sound pressure level (Leq), which is expressed in units ofdBA and corresponds to fictitious noise source emitting steady noise,which in a specific period of time contains the same acoustic energyas the observed source with fluctuating noise. The Leq for time intervalbetween times t1 and t2 in seconds is expressed by the followingequation:

Leq ¼ 10 log1

t2−t1

Z t2

t1

p2Ap20

dt

" #ð1Þ

where pA(Pa) is the time varying sound pressure and p0 is a referencesound pressure taken as 20 μPa.

In order to predict the noise it is necessary to know the functionalrelationship between the equivalent sound pressure level and the influ-ential parameters. Leq is correlated to numerous parameters, such asnumbers and types of vehicles, their velocities, type of road surface,width and slope of the road, and height of buildings facing the road.

Table 1Definition of acoustic classes of vehicles.

Acoustic class QLTC-10C classification

LMV A1, A2, B1STV B2, B3TTV B4, B5BUS C1, C2

A0, XX

As mentioned in the Introduction section, in this paper the followingvariables were considered: the number of light motor vehicles (LMV),the number of medium trucks (STV), the number of heavy trucks(TTV), the number of buses (BUS) and the average traffic flow speed(Vavg). A brief description of how these variables were measured isgiven in the following section.

Data measurement procedure

The research was carried out on a two-lane motorway road with atwo-way traffic. The number of vehicles per particular acoustic classand the average traffic flow speed were determined by means ofautomatic vehicle counter QLTC-10C. This appliance operates with twoinductive loopsmounted onto the road surface, whichmakes it possibleto classify vehicles and calculate the average traffic flow speed. Itrecognizes 11 subclasses of vehicles in compliance with the EEC 1108/70 EU Directive (http://www.mikrobit.si/pages/eng/hardware/QLTC-10.htm). On the basis of that classification we defined four acousticclasses: light motor vehicles (LMV), medium trucks (STV), heavy trucks(TTV), and buses (BUS), by uniting some subclasses as presented inTable 1.

Subclasses A0 and XX were not taken in consideration because of avery small number of vehicles of this type in the analyzed traffic flow.

For measuring the traffic noise the noise level meter Bruel & Kajertype 2230 whose measurement error is 0.1 dB was used. The noisedetectionwas carried out in the Fast regime. The noise level ismeasured

Fig. 2. Information processing in ANN.

Fig. 3. Structure of proposed ANN for Leq prediction.

26 V. Nedic et al. / Environmental Impact Assessment Review 49 (2014) 24–30

as the equivalent level (Leq) at A weight scale, expressed in dBA for aperiod of 1 h. In order to include a greater number of scenarios thatmight occur in urban environments, a total of 120 measurements ofequivalent noise levels were carried out. Measurements of Leq were

Fig. 4. ANN s

performed at various times to include diversity of the traffic flow asmuch as possible. Simultaneously, variations in traffic flow, trafficspeed and composition of traffic flow were measured. For that reasonthe surveys at the same time also consist of the following parameters:

oftware.

Table 2Neural network weights (NP = neurons in previous layer; B = bias).

Neurons in hidden layer 1

NP 1 2 3 4 5 6 7 8

1 −5.36312 4.46471 0.97817 −0.15320 4.77468 0.92228 5.10674 −8.203682 2.17799 −8.71264 0.86297 1.12126 −1.37285 0.14763 −6.51881 −10.184043 −5.81881 −2.42493 −7.18756 2.14258 4.56233 2.60649 1.68611 −4.931184 1.38220 −7.01484 −3.26992 −1.66404 −2.46716 2.97222 1.36237 0.555185 −2.93764 −0.18404 −9.99148 −3.12384 −0.93832 3.05930 5.10282 −0.80884B 2.28816 1.45341 2.00800 −1.57072 −1.97450 0.31079 −5.64566 8.37095

Neurons in hidden layer 2

NP 1 2 3

1 −2.61995 −5.45037 1.707092 −2.40335 −7.31199 −1.849593 −6.41376 2.91301 3.259674 2.98875 0.03611 0.047135 −2.34561 −0.01629 −5.549226 1.76475 −1.61923 −4.119717 5.35269 −1.76336 −2.165828 −4.86500 4.48768 5.77036B −1.38983 −0.69197 −4.23090

Neurons in output layerNP 1

1 5.856042 5.138633 −6.17416B −0.63755

0 20 40 60 8067

68

69

70

71

72

73

74

Sample

Leq

(dB

A)

ExperimentalPredicted

Fig. 5. Side by side comparison of measured data and calculation for learning process.

27V. Nedic et al. / Environmental Impact Assessment Review 49 (2014) 24–30

the number of light motor vehicles, the number of medium trucks, thenumber of heavy trucks, the number of buses, and the average trafficspeed in the given time periods.

Measurementswere taken in accordancewith recommendations forroad traffic noisemeasurement; amicrophonewasmounted away fromreflecting facades, at a height of 1.2 m above the ground level and 7.5 maway from the central line of the road. During the measurements it hasbeen taken care that climate conditions were as similar as possible (nowind, no rain) in order to reduce their influence.

Neural network model

An ANN is a computational model based on biological neuralnetworks. It consists of interconnected artificial neurons that aregrouped in input, hidden and output layers (Fig. 1).

The way in which ANN processes information from one layer toother is presented in Fig. 2 and could be mathematically depicted bythe following two equations:

vk ¼Xni¼1

xiwi þ bk ð2Þ

yk ¼ φ vkð Þ ð3Þ

where bk is bias, and φ is activation function and wi are weights.The activation function must be differentiable, and its choice

depends on implementation. Influence of different activation functionson neural network performance was investigated in the work of da S.Gomes et al. (2010), Karlik and Olgac (2010) and Nait Chairf Hammadi(1998). In this work the sigmoid function, which is defined by thefollowing equation is used:

φ xð Þ ¼ 11þ e−x : ð4Þ

The number of input and output neurons in neural network is deter-mined by the nature of the problem. The number of hidden neurons, aswell as the number of hidden layers, has in general a significant influ-ence on the final output. Neural networks with more hidden layerscanmodel complex relationships between independent and dependentvariables with any kind of shape. Apart from some rules of thumb thereis no general algorithm for determining the optimal number of neuronsin the hidden layers. This number is determined empirically for a partic-ular instance. If there are too few neurons in the hidden layers, thenetwork could not properly catch the signals from the input neurons.On the other hand, too many neurons in the hidden layers can resultin overfitting.

0 0.5 10

0.5

1

y=1.0084x+−0.0004R2=0.9305

T

A

A=TData pointsLinear (Data points)

Fig. 8. Comparison of measured data (T) and calculation (A) for test data set.

0 0.5 10

0.5

1

y=0.9706x+0.0381R2=0.9779

T

A

A=TData pointsLinear (Data points)

Fig. 6. Comparison of measured data (T) and calculation (A) for learning process.

28 V. Nedic et al. / Environmental Impact Assessment Review 49 (2014) 24–30

Simulation arrangement

The neural network architecture that is used in this work is made upof one output neuron corresponding to the value of Leq, two hiddenlayers with eight and three neurons, and five inputs referring respec-tively to: the number of light motor vehicles, the number of mediumtrucks, the number of heavy trucks, the number of buses and theaverage traffic flow speed (Fig. 3).

The simulation was done using a user friendly software developedby the authors, in the VB 6.0 framework (Fig. 4).

In order to have an estimation of predictive ability of themodel withdata that were not used in the learning process, the measurement dataset is divided into two subsets: training (learning) data set and test dataset. In this work training and test data set consist of 80 and 40 samplesrespectively. Dividing the data set into learning and test subsets has asignificant impact on the performance of the ANN model (WenyanWu et al., 2012). The conventional approach that is most oftenemployed in the literature involves arbitrarily dividing the data(Bowden et al., 2002). In this work the subsets are formed by randomselection of measurements, taking care that test data do not fall outside

0 10 20 30 4067

68

69

70

71

72

73

74

Sample

Leq

(dB

A)

ExperimentalPredicted

Fig. 7. Side by side comparison of measured data and calculation for test data set.

of the range of the data used for training. This is done automatically bythe software, but software also offers a possibility to separately inputtraining and test data, so these subsets can be created externally.

The dendrites' weight values of the corresponding neural networkarchitecture that are obtained after training the network on the trainingdata set are presented in Table 2.

Results

Side by side comparison of measured values and computationalresults for learning data set is shown in Fig. 5 and this comparison inregression scatter plot is presented in Fig. 6. Side by side comparison ofmeasured values and computational results for test data set is shownin Fig. 7 and this comparison in regression scatter plot is presented inFig. 8. These pictures clearly demonstrate that there is relatively goodagreement between measured data and computational results for bothdata sets.

0 10 20 30 40Sample

56

60

64

68

72

dBA

MeasurementANN

BurgessGriffiths

FagottiRLS 90

C.S.T.B.C.N.R.

Fig. 9. Side by side comparison of ANN model and various statistical models.

Table 3Statistical analysis for different calculation methods for test data set.

Model Min. Max. Average St. error St. deviation R R2 ME MAE MAPE F value t value

Measurement 67.4 73.4 71.31 0.205 1.299 – –

[gray] 0.8 ANN model 67.76 73.64 71.34 0.215 1.357 0.965 0.931 0.03 0.301 0.004 1.093 −0.101Burgess 58.72 64.21 62.1 0.175 1.107 0.91 0.828 −9.213 9.213 0.148 1.375 34.143Griffiths 60.91 66.35 64.83 0.153 0.967 0.87 0.758 −6.481 6.481 0.1 1.804 25.318Fagotti 55.54 63.5 59.41 0.253 1.601 0.67 0.449 −11.906 11.906 0.201 1.519 36.534RLS 90 56.44 62.13 60.3 0.17 1.075 0.9 0.81 −11.008 11.008 0.183 1.458 41.291C.S.T.B. 61.69 66.76 65.89 0.13 0.822 0.667 0.445 −5.425 5.425 0.082 2.497 22.327C.N.R. 59.75 65.76 64.14 0.162 1.026 0.854 0.729 −7.17 7.17 0.112 1.603 27.405

tpk — table = 1.96 (p = 5%, k = 78).F k1 ;k2p — table = 1.52 (p = 5%, k1 = 39, k2 = 39).

29V. Nedic et al. / Environmental Impact Assessment Review 49 (2014) 24–30

The neural network based approach versus the classical approach

In order to assess the advantages of the ANN model in traffic noiseprediction, the simulation results are comparedwith the results obtainedusing classical methods. Thesemethodswere reviewed in Quartieri et al.(2009) and are reported bellow:

Burgess : Leq ¼ 55:5þ 10:2 log Qð Þ þ 0:3P−19:3 log dð ÞFagotti : Leq ¼ 10 log QL þ QM þ 8QP þ 88QBUSð Þ þ 33:5

Griffiths : Leq ¼ L50 þ 0:018 L10−L90ð Þ2L10 ¼ 61þ 8:4 log Qð Þ þ 0:15P−11:5 log dð ÞL50 ¼ 44:8þ 10:8 log Qð Þ þ 0:12P−9:6 log dð ÞL90 ¼ 39:1þ 10:5 log Qð Þ þ 0:06P−9:3 log dð Þ

C:S:T:B: : Leq ¼ 0:65L50 þ 28:8L50 ¼ 11:9 log Qð Þ þ 31:4

RLS 90 : Lm;E 25ð Þ ¼ 37:3þ 10 log Q 1þ 0:082Pð Þ½ �C:N:R: : Leq ¼ 35:1þ 10 log QL þ 6QPð Þ−10 log d=25ð Þ þ 1:5

Here Q is the number of vehicles per hour, P is the percentage ofheavy vehicles, d is the distance from observation point to the centerof the traffic lane, QL is the number of light vehicles per hour, QP is thenumber of heavy vehicles per hour, QM is the number of motorcyclesper hour, QBUS is the number of buses per hour, and L is the width ofthe road. The Lm,E(25), which is given in the RLS 90 model pertains tothe distance of 25 m from a measurable point to the center of the roadline, and is transformed to the distance of 7.5mby adding the correctionterm 10 log(25/7.5).

The results of the comparison are given in Fig. 9 and Table 3.Predicting capabilities are compared using various statistical parame-ters: standard error, standard deviation, mean error (ME), mean abso-lute error (MAE), mean absolute percentage error (MAPE), coefficientof correlation (R) and coefficient of determination (R2) (Kennedy,2003). Apart from that, t-test and F-test are used for statistical analysisand comparison of results of different models with measured values.The null hypothesis of the F-test shows that the dispersion of the mea-sured noise values and values calculated by neural network model areapproximately equal. For the degrees of freedom k1 = 39, k2 = 39 andp = 0.05 the calculated F value for neural network model is less thanthe table value of F0.0539,39 = 1.52 (Table 3). This leads to a conclusionthat the dispersion of the measured and calculated values does not dif-fer significantly, and there is no basis for rejecting the null hypothesis ofequal dispersion of data sets. This is not the case with the Griffiths,C.S.T.B. and C.N.R. models.

Moreover, t-test shows that themeanvalue of themeasured and cal-culated noise values are approximately equal when neural networkmodel is concerned. For the degree of freedom k = 78 and p = 0.05the calculated t value is less than the table value t0.05

78 = 1.96(Table 3), and there is no reason to reject the null hypothesis of equalityof means of data sets. This is not the case with all of analyzed statisticalmodels where the calculated t value is significantly more than the tablevalue and the null hypothesis of equality of means of the data sets mustbe rejected.

Other statistical parameters that are shown in Table 3 also demon-strate better prediction capabilities of neural network model whencompared with analyzed empirical relationships, and this is also obviousin Fig. 9. This advantage of neural networks is due to their greater capacityin approximating non-linear relationship between the traffic flowstructure and the equivalent noise level.

Conclusions

Noise pollution near urban arterials has a complex relationship withmany influential factors. These relationships are highly non-linear andnot well modeled with classical statistical methods. In this paper factorsthat affect the noise level are divided into five groups and all have beenused as inputs in neural networks to predict equivalent noise level Leq.The neural network is designed, learned on randomly selected data ex-tracted from the measurement data set and tested using originally de-veloped user friendly software package. Predictive capabilities ofdeveloped ANN model are assessed by various statistical parameters.Furthermore, the results are compared with outputs of some classicalmodels, and it could be concluded that artificial neural networks showmuch better capabilities to predict equivalent noise level based on thetraffic flow structure than any other statistical method. The advantageof the proposed model compared to the ANN model presented inCammarata et al. (1995) is that it is not sensitive to town layout.

References

BabischW, Pershagen G, Selander J, Houthuijs D, Breugelmans O, Cadum E, et al. Noise an-noyance — a modifier of the association between noise level and cardiovascularhealth? Sci Total Environ 2013;452–453:50–7.

Bowden GJ, Maier HR, Dandy GC. Optimal division of data for neural network models inwater resources applications. Water Resour Res 2002;38(2):2-1–2-11.

Brink M. Parameters of well-being and subjective health and their relationship with res-idential traffic noise exposure — a representative evaluation in Switzerland. EnvironInt 2011;37(4):723–33.

Caciari T, Rosati MV, Casale T, Loreti B, Sancini A, Riservato R, et al. Noise-induced hearingloss in workers exposed to urban stressors. Sci Total Environ 2013;463–464:302–8.

Calixto A, Diniz FB, Zannin PH. The statistical modeling of road traffic noise in an urbansetting. Cities 2003;20(1):23–9.

Cammarata G, Cavalieri S, Fichera A. A neural network architecture for noise prediction.Neural Netw 1995;8(6):963–73.

da S. Gomes GS, Ludermir TB, Lima LMMR. Comparison of new activation functions inneural network for forecasting financial time series, 20; 2010. p. 417–39.

Fyhri A, Klboe R. Road traffic noise, sensitivity, annoyance and self-reported health — astructural equation model exercise. Environ Int 2009;35(1):91–7.

Garg N, Maji S. A critical review of principal traffic noise models: strategies and implica-tions. Environ Impact Assess Rev 2014;46:68–81.

Givargis S, Karimi H. A basic neural traffic noise prediction model for Tehran's roads.J Environ Manag 2010;91(12):2529–34.

Guarnaccia C, Lenza TLL, Mastorakis NE, Quartieri J. A comparison between traffic noiseexperimental data and predictive models results. Int J Mech 2011;5(4):379–86.

Güdogdu Ö, Gökdag M, Yüksel F. A traffic noise prediction method based on vehiclecomposition using genetic algorithms. Appl Acoust 2005;66(7):799–809.

Karlik B, Olgac AV. Performance analysis of various activation functions in generalizedmlp architectures of neural networks, 1; 2010. p. 111–22.

Kassomenos P, Vogiatzis K, Coelho JLB. Critical issues on environmental noise: editorial.Sci Total Environ 2014;399:482–3. http://dx.doi.org/10.1016/j.scitotenv.2014.02.108.

Kennedy P. A guide to econometrics. Blackwell; 2003.Nait Chairf Hammadi HI. On the activation function and fault tolerance in feed forward

neural networks, E81-D; 1998. p. 66–72.

30 V. Nedic et al. / Environmental Impact Assessment Review 49 (2014) 24–30

Pirrera S, De Valck E, Cluydts R. Nocturnal road traffic noise: a review on its assessmentand consequences on sleep and health. Environ Int 2010;36(5):492–8.

Quartieri J, Mastorakis N, Iannone G, Guarnaccia C, DAmbrosio S, Troisi A, et al. A review oftraffic noise predictive models. The 5th WSEAS International Conference on Appliedand Theoretical Mechanics. Puerto De La Cruz, Canary Islands; 2009.

Rahmani S, Mousavi SM, Kamali MJ. Modeling of road-traffic noise with the use of geneticalgorithm. Appl Soft Comput 2011;11(1):1008–13.

Steele C. A critical review of some traffic noise prediction models. Appl Acoust 2001;62(3):271–87.

Slobodan Cvetanovic is Full Professor at the Faculty ofEconomics, University of Nis, Serbia. His current researchinterest includes macroeconomics, theory and policy ofeconomic development, innovation policy, and environ-mental and natural resource economics.

VladimirNedic is ITmanager at the Faculty of Philology andArts at the University of Kragujevac, Serbia. He is currently aPhD student at the Faculty of Engineering at the Universityof Kragujevac, Serbia. His key scientific interests are: qualitysystem management, software support of system manage-ment, organizational changes, and document management.

Danijela Despotovic is Assistant Professor at the Facultyof Economics, University of Kragujevac, Serbia. Her currentresearch interest includes macroeconomic theory and poli-cy, sustainable development, innovation policy, environ-mental policy, and optimization.

Suarez E, Barros J. Traffic noise mapping of the city of Santiago de Chile. Sci Total Environ2014;466–467:539–46.

WuW, Maier HR, Dandy GC, May R. Exploring the impact of data splitting methods on ar-tificial neural network models. 10th International Conference on Hydroinformatics,HIC 2012, Hamburg, Germany; 2012.

Milan Despotovic is Full Professor at the Department of En-ergy and Process Engineering at the Faculty of Engineering,University of Kragujevac, Serbia. His current research inter-est includes energy management, renewable energysources, optimization and artificial intelligence.

Sasa Babic holds a PhD degree at the Faculty of Engineering,University of Kragujevac, Serbia. He is currently a Research As-sistant at the College of Applied Mechanical Engineering,Trstenik, Serbia. His current research interests are trafficmanagement and traffic noise.