Artiﬁcial Neural Networks and Fuzzy Neural Networks for ...

Complexity

Artificial Neural Networks and Fuzzy Neural Networks for Solving Civil Engineering Problems

Lead Guest Editor Milos KnezevicGuest Editors Meri Cvetkovska Tomaacuteš Hanaacutek Luis Braganca and Andrej Soltesz

Artificial Neural Networks and Fuzzy NeuralNetworks for Solving Civil Engineering Problems

Complexity


Lead Guest Editor Milos KnezevicGuest Editors Meri Cvetkovska Tomaacuteš Hanaacutek Luis Bragancaand Andrej Soltesz

Copyright copy 2018 Hindawi All rights reserved

This is a special issue published in ldquoComplexityrdquo All articles are open access articles distributed under the Creative Commons Attribu-tion License which permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Editorial Board

Joseacute A Acosta SpainCarlos F Aguilar-Ibaacutentildeez MexicoTarek Ahmed-Ali FranceBasil M Al-Hadithi SpainJuan A Almendral SpainDiego R Amancio BrazilDavid Arroyo SpainMohamed Boutayeb FranceArturo Buscarino ItalyGuido Caldarelli ItalyEric Campos-Canton MexicoMohammed Chadli FranceDiyi Chen ChinaGiulio Cimini ItalyDanilo Comminiello ItalySara Dadras USAManlio De Domenico ItalyPietro De Lellis ItalyAlbert Diaz-Guilera SpainThach Ngoc Dinh FranceJordi Duch SpainMarcio Eisencraft BrazilJoshua Epstein USAMondher Farza FranceThierry Floquet FranceMattia Frasca ItalyLucia Valentina Gambuzza ItalyBernhard C Geiger Austria

Carlos Gershenson MexicoPeter Giesl UKSergio Goacutemez SpainLingzhong Guo UKXianggui Guo ChinaSigurdur F Hafstein IcelandChittaranjan Hens IsraelGiacomo Innocenti ItalySarangapani Jagannathan USAMahdi Jalili AustraliaJeffrey H Johnson UKM Hassan Khooban DenmarkVincent Labatut FranceLucas Lacasa UKQingdu Li GermanyChongyang Liu ChinaXiaoping Liu CanadaR M Lopez Gutierrez MexicoVittorio Loreto ItalyDidier Maquin FranceEulalia Martiacutenez SpainMarcelo Messias BrazilAna Meštrović CroatiaLudovico Minati JapanCh P Monterola PhilippinesMarcin Mrugalski PolandRoberto Natella ItalyNam-Phong Nguyen USA

Beatrice M Ombuki-Berman CanadaIrene Otero-Muras SpainYongping Pan SingaporeDaniela Paolotti ItalyCornelio Posadas-Castillo MexicoMahardhika Pratama SingaporeLuis M Rocha USAMiguel Romance SpainAvimanyu Sahoo USAMatilde Santos SpainHiroki Sayama USAMichele Scarpiniti ItalyEnzo Pasquale Scilingo ItalyDan Selişteanu RomaniaDimitrios Stamovlasis GreeceSamuel Stanton USARoberto Tonelli ItalyShahadat Uddin AustraliaGaetano Valenza ItalyDimitri Volchenkov USAChristos Volos GreeceZidong Wang UKYan-Ling Wei SingaporeHonglei Xu AustraliaXinggang Yan UKMassimiliano Zanin SpainHassan Zargarzadeh USARongqing Zhang USA

Contents

Artificial Neural Networks and Fuzzy Neural Networks for Solving Civil Engineering ProblemsMilos Knezevic Meri Cvetkovska Tomaacuteš Hanaacutek Luis Braganca and Andrej SolteszEditorial (2 pages) Article ID 8149650 Volume 2018 (2018)

Determination of Fire Resistance of Eccentrically Loaded Reinforced Concrete Columns Using FuzzyNeural NetworksMarijana Lazarevska Ana Trombeva Gavriloska Mirjana Laban Milos Knezevic and Meri CvetkovskaResearch Article (12 pages) Article ID 8204568 Volume 2018 (2018)

Urban Road Infrastructure Maintenance Planning with Application of Neural NetworksIvan Marović Ivica Androjić Nikša Jajac and Tomaacuteš HanaacutekResearch Article (10 pages) Article ID 5160417 Volume 2018 (2018)

ANN Based Approach for Estimation of Construction Costs of Sports FieldsMichał Juszczyk Agnieszka Leśniak and Krzysztof ZimaResearch Article (11 pages) Article ID 7952434 Volume 2018 (2018)

Assessment of the Real Estate Market Value in the European Market by Artificial Neural NetworksApplicationJasmina Ćetković Slobodan Lakić Marijana Lazarevska Miloš Žarković Saša Vujošević Jelena Cvijovićand Mladen GogićResearch Article (10 pages) Article ID 1472957 Volume 2018 (2018)

Development of ANNModel for Wind Speed Prediction as a Support for Early Warning SystemIvan Marović Ivana Sušanj and Nevenka OžanićResearch Article (10 pages) Article ID 3418145 Volume 2017 (2018)

Estimation of Costs and Durations of Construction of Urban Roads Using ANN and SVMIgor Peško Vladimir Mučenski Miloš Šešlija Nebojša Radović Aleksandra Vujkov Dragana Bibićand Milena KrklješResearch Article (13 pages) Article ID 2450370 Volume 2017 (2018)

EditorialArtificial Neural Networks and Fuzzy Neural Networks for SolvingCivil Engineering Problems

Milos Knezevic 1 Meri Cvetkovska 2 Tomaacuteš Hanaacutek 3 Luis Braganca 4

and Andrej Soltesz5

1University of Podgorica Faculty of Civil Engineering Podgorica Montenegro2Ss Cyril and Methodius University Faculty of Civil Engineering Skopje Macedonia3Brno University of Technology Faculty of Civil Engineering Institute of Structural Economics and ManagementBrno Czech Republic4Director of the Building Physics amp Construction Technology Laboratory Civil Engineering Department University of MinhoGuimaraes Portugal5Slovak University of Technology in Bratislava Department of Hydraulic Engineering Bratislava Slovakia

Correspondence should be addressed to Milos Knezevic knezevicmiloshotmailcom

Received 2 August 2018 Accepted 2 August 2018 Published 8 October 2018

Copyright copy 2018 Milos Knezevic et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Based on the live cycle engineering aspects such as predictiondesign assessment maintenance and management of struc-tures and according to performance-based approach civilengineering structures have to fulfill essential requirementsfor resilience sustainability and safety from possible risks suchas earthquakes fires floods extreme winds and explosions

The analysis of the performance indicators which are ofgreat importance for the structural behavior and for the fulfill-ment of the above-mentioned requirements is impossiblewithout conducting complex mathematical calculations Arti-ficial neural networks and Fuzzy neural networks are typicalexamples of a modern interdisciplinary field which gives thebasic knowledge principles that could be used for solvingmany different and complex engineering problems whichcould not be solved otherwise (using traditional modelingand statistical methods) Neural networks are capable of col-lecting memorizing analyzing and processing a large numberof data gained from some experiments or numerical analysesBecause of that neural networks are often better calculationand prediction methods compared to some of the classicaland traditional calculation methods They are excellent in pre-dicting data and they can be used for creating prognosticmodels that could solve various engineering problems andtasks A trained neural network serves as an analytical toolfor qualified prognoses of the results for any input data which

have not been included in the learning process of the networkTheir usage is reasonably simple and easy yet correct and pre-cise These positive effects completely justify their applicationas prognostic models in engineering researches

The objective of this special issue was to highlight thepossibilities of using artificial neural networks and fuzzyneural networks as effective and powerful tools for solvingengineering problems From 12 submissions 6 papers arepublished Each paper was reviewed by at least two reviewersand revised according to review comments The papers cov-ered a wide range of topics such as assessment of the realestate market value estimation of costs and duration of con-struction works as well as maintenance costs and predictionof natural disasters such as wind and fire and prediction ofdamages to property and the environment

I Marovic et alrsquos paper presents an application of arti-ficial neural networks (ANN) in the predicting process ofwind speed and its implementation in early warning sys-tems (EWS) as a decision support tool The ANN predic-tion model was developed on the basis of the input dataobtained by the local meteorological station The predic-tion model was validated and evaluated by visual andcommon calculation approaches after which it was foundout that it is applicable and gives very good wind speedpredictions The developed model is implemented in the

HindawiComplexityVolume 2018 Article ID 8149650 2 pageshttpsdoiorg10115520188149650

EWS as a decision support for the improvement of theexisting ldquoprocedure plan in a case of the emergency causedby stormy wind or hurricane snow and occurrence of theice on the University of Rijeka campusrdquo

The application of artificial neural networks as well aseconometric models is characterized by specific advantagesand disadvantages Nevertheless neural networks have beenimposed as a real alternative to econometric methods andas a powerful tool for assessment and forecasting for exam-ple in the field of evaluating real estate It is specially empha-sized that it is possible to find estimated values instead ofexact values The aim of J Cetkovic et alrsquos research was toconstruct a prognostic model of the real estate market valuein the EU countries depending on the impact of macroeco-nomic indicators Based on the available input datamdashma-croeconomic variables that influence the determination ofreal estate prices the authors sought to obtain fairly correctoutput datamdashprices forecast in the real estate markets ofthe observed countries

Offer preparation has always been a specific part of abuilding process which has a significant impact on companybusiness Due to the fact that income greatly depends onofferrsquos precision and the balance between planned costs bothdirect and overheads and wished profit it is necessary to pre-pare a precise offer within the required time and availableresources which are always insufficient I Peško et alrsquos paperpresents research on precision that can be achieved whileusing artificial intelligence for the estimation of cost andduration in construction projects Both artificial neural net-works (ANNs) and support vector machines (SVM) wereanalyzed and compared Based on the investigation resultsa conclusion was drawn that a greater accuracy level in theestimation of costs and duration of construction is achievedby using models that separately estimate the costs and theduration The reason for this lies primarily in the differentinfluence of input parameters on the estimation of costs incomparison with the estimation of duration of the projectBy integrating them into a single model a compromise interms of the significance of input data is made resulting inthe lower precision of estimation when it comes to ANNmodels SVMmodels feature a greater capacity of generaliza-tion providing at the same time greater accuracy of estima-tion both for the estimation of costs and duration ofprojects as well

The same problem was treated by M Juszczyk et alTheir research was on the applicability of ANN for theestimation of construction costs of sports fields An appli-cability of multilayer perceptron networks was confirmedby the results of the initial training of a set of various arti-ficial neural networks Moreover one network was tailoredfor mapping a relationship between the total cost of con-struction works and the selected cost predictors whichare characteristic for sports fields Its prediction qualityand accuracy were assessed positively The research resultslegitimate the proposed approach

The maintenance planning within the urban road infra-structure management is a complex problem from both themanagement and the technoeconomic aspects The focus ofI Marovic et alrsquos research was on decision-making processes

related to the planning phase during the management ofurban road infrastructure projects The goal of thisresearch was to design and develop an ANN model inorder to achieve a successful prediction of road deteriora-tion as a tool for maintenance planning activities Such amodel was part of the proposed decision support conceptfor urban road infrastructure management and a decisionsupport tool in planning activities The input data wereobtained from Circly 60 Pavement Design Software andused to determine the stress values It was found that itis possible and desirable to apply such a model in thedecision support concept in order to improve urban roadinfrastructure maintenance planning processes

The fire resistance of civil engineering structures can bedetermined based on the estimated fire resistance of eachconstruction element (columns beams slabs walls etc)As fire resistance of structural elements directly affects thefunctionality and safety of the whole structure the signifi-cance which new methods and computational tools have onenabling a quick easy and simple prognosis of the same isquite clear M Lazarevska et alrsquos paper considered the appli-cation of fuzzy neural networks by creating prognosticmodels for determining fire resistance of eccentrically loadedreinforced concrete columns Using the concept of the fuzzyneural networks and the results of the performed numericalanalyses (as input parameters) the prediction model fordefining the fire resistance of eccentrically loaded RC col-umns incorporated in walls and exposed to standard firefrom one side has been made The numerical results wereused as input data in order to create and train the fuzzy neu-ral network so it can provide precise outputs for the fire resis-tance of eccentrically loaded RC columns for any other inputdata (RC columns with different dimensions of the cross-sec-tion different thickness of the protective concrete layer dif-ferent percentage of reinforcement and for different loads)

These papers represent an exciting insightful observationinto the state of the art as well as emerging future topics inthis important interdisciplinary field We hope that this spe-cial issue would attract a major attention of the civil engi-neeringrsquos community

We would like to express our appreciation to all theauthors and reviewers who contributed to publishing thisspecial issue

Conflicts of Interest

As guest editors we declare that we do not have a financialinterest regarding the publication of this special issue

Milos KnezevicMeri CvetkovskaTomaacuteš HanaacutekLuis BragancaAndrej Soltesz

2 Complexity

Research ArticleDetermination of Fire Resistance of Eccentrically LoadedReinforced Concrete Columns Using Fuzzy Neural Networks

Marijana Lazarevska 1 Ana Trombeva Gavriloska2 Mirjana Laban3 Milos Knezevic 4

and Meri Cvetkovska1

1Faculty of Civil Engineering University Ss Cyril and Methodius 1000 Skopje Macedonia2Faculty of Architecture University Ss Cyril and Methodius 1000 Skopje Macedonia3Faculty of Technical Sciences University of Novi Sad Novi Sad Serbia4Faculty of Civil Engineering University of Montenegro 81000 Podgorica Montenegro

Correspondence should be addressed to Marijana Lazarevska marijanagfukimedumk

Received 21 February 2018 Accepted 8 July 2018 Published 23 August 2018

Academic Editor Matilde Santos

Copyright copy 2018Marijana Lazarevska et al This is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

Artificial neural networks in interaction with fuzzy logic genetic algorithms and fuzzy neural networks represent an example of amodern interdisciplinary field especially when it comes to solving certain types of engineering problems that could not be solvedusing traditional modeling methods and statistical methods They represent a modern trend in practical developments within theprognostic modeling field and with acceptable limitations enjoy a generally recognized perspective for application in constructionResults obtained from numerical analysis which includes analysis of the behavior of reinforced concrete elements and linearstructures exposed to actions of standard fire were used for the development of a prognostic model with the application of fuzzyneural networks As fire resistance directly affects the functionality and safety of structures the significance which new methodsand computational tools have on enabling quick easy and simple prognosis of the same is quite clear This paper will considerthe application of fuzzy neural networks by creating prognostic models for determining fire resistance of eccentrically loadedreinforced concrete columns

1 Introduction

The fire resistance of civil engineering structures can bedetermined based on the estimated fire resistance of eachconstruction element (columns beams slabs walls etc)The fire resistance of a structural element is the time period(in minutes) from the start of the fire until the moment whenthe element reaches its ultimate capacity (ultimate strengthstability and deformability) or until the element loses theinsulation and its separation function [1] The legallyprescribed values for the fire resistance can be achieved byapplication of various measures (by using appropriate shapeand elementrsquos dimensions and proper static system thermo-isolation etc) The type of applied protection measuresmainly depend on the type of construction material thatneeds to be protected Different construction materials

(concrete steel and wood) have different behaviors underelevated temperatures That is why they have to be protectedin accordance with their individual characteristics whenexposed to fire [1] Even though the legally prescribed valuesof the fire resistance is of huge importance for the safety ofevery engineering structures in Macedonia there is noexplicit legally binding regulation for the fire resistanceThe official national codes in the Republic of Macedoniaare not being upgraded and the establishment of new codesis still a continuing process Furthermore most of the exper-imental models for determination of fire resistance areextremely expensive and analytical models are quite compli-cated and time-consuming A modern type of analyses suchas modeling through neural networks can be very helpfulparticularly in those cases where some prior analyses werealready made Therefore the application of artificial and


fuzzy neural networks for prognostic modeling of the fireresistance of structures is of significant importance especiallyduring the design phase of civil engineering structures

Fuzzy neural networks are typical example of a moderninterdisciplinary subject that helps solving different engi-neering problems which cannot be solved by the traditionalmodeling methods [2ndash4] They are capable of collectingmemorizing analyzing and processing large number of dataobtained from some experiments or numerical analyses Thetrained fuzzy neural network serves as an analytical tool forprecise predictions for any input data which are not includedin the training or testing process of the model Their opera-tion is reasonably simple and easy yet correct and precise

Using the concept of the fuzzy neural networks and theresults of the performed numerical analyses (as input param-eters) the prediction model for defining the fire resistance ofeccentrically loaded RC columns incorporated in walls andexposed to standard fire from one side has been made

The goal of the research presented in this paper was tobuild a prognostic model which could generate outputs forthe fire resistance of RC columns incorporated in walls forany given input data by using the results from the conductednumerical analyses The numerical results were used as inputdata in order to create and train the fuzzy neural network soit can provide precise outputs for the fire resistance ofeccentrically loaded RC columns for any other input data(RC columns with different dimensions of the cross sectiondifferent thickness of the protective concrete layer differentpercentage of reinforcement and for different loads)

2 Fuzzy Neural Networks Theoretical Basis

Fuzzy neural networks are defined as a combination ofartificial neural networks and fuzzy systems in such a waythat learning algorithms from neural networks are used todetermine the parameters of a fuzzy system One of the mostimportant aspects of this combination is that the system canalways be interpreted using the ldquoif-thenrdquo rule because it isbased on a fuzzy system that reflects uncertainunclearknowledge Fuzzy neural networks use linguistic knowledgefrom the fuzzy system and learning ability from neuralnetworks Therefore fuzzy neural networks are capable ofprecisely modeling ambiguity imprecision and uncertaintyof data with the additional learning opportunity characteris-tic of neural networks [3 5ndash8]

Fuzzy neural networks are based on a common conceptof fuzzy logic and artificial neural networks theories thatare already at the top of the list for researchers of artificialintelligence Fuzzy logic based on Zadehrsquos principle of fuzzysets provides mathematical potential for describing theuncertainty that is associated with cognitive processes ofthinking and reasoning This makes it possible to drawconclusions even with incomplete and insufficiently preciseinformation (so-called approximate conclusions) On theother hand artificial neural networks with their variousarchitectures built on the artificial neuron concept have beendeveloped as an imitation of the biological neural system forthe successful performance of learning and recognitionfunctions What is expected from the fusion of these two

structures is that the learning and computational ability ofneural networks will be transmitted into the fuzzy systemand that the highly productive if-then thinking of the fuzzysystem will be transferred to neural networks This wouldallow neural networks to be more than simply ldquoblack boxesrdquowhile fuzzy inference systems will be given the opportunity toautomatically adjust their parameters [2 3 5ndash8]

Depending on the field of application several approacheshave been developed for connecting artificial neural networksand fuzzy inference systems which are most often classifiedinto the following three groups [3 5 7ndash9] cooperativemodels concurrent models and integrated (hybrid) models

The basic characteristic of the cooperative model is thatvia learning mechanisms of artificial neural networksparameters of the fuzzy inference system are determinedthrough training data which allows for its quick adaptionto the problem at hand A neural network is used todetermine the membership function of the fuzzy systemthe parameters for fuzzy rules weight coefficients and othernecessary parameters Fuzzy rules are usually determinedusing clustering access (self-organizing) while membershipfunctions are elicited from training data using a neuralnetwork [3 5 7ndash9]

Characteristic of the concurrent model is that the neuralnetwork continuously assists the fuzzy inference systemduring the process of determining and adjusting requiredparameters In some cases the neural network can correctoutput results while in other cases it corrects input data intothe fuzzy inference system [3 5 7ndash9]

For integrated fuzzy neural networks the learningalgorithm from a neural network is used to determine theparameters of the fuzzy inference system These networksrepresent a modern class of fuzzy neural networks character-ized by a homogeneous structure that is they can beunderstood as neural networks represented by fuzzy param-eters [3 5ndash9] Different models for hybrid fuzzy neuralnetworks have been developed among which the followingstand out FALCON ANFIS GARIC NEFCON FUNSONFIN FINEST and EFuNN

3 State-of-the-Art Application of FuzzyNeural Networks

In the civil engineering field fuzzy neural networks are veryoften used to predict the behavior of materials and con-structive elements The main goal of such prognostic modelsis to obtain a solution to a problem by prediction (mappinginput variables into corresponding output values) For thequalitative development of efficient prognostic models it isnecessary to have a number of data groups Fortunatelywhen it comes to civil engineering data collection is not amajor problem which enhances the possibility of applyingsuch innovative techniques and methods Some examples ofsuccessful application of fuzzy neural networks to variousfields of civil engineering are presented in the followingsection of this paper [4 5]

Fuzzy neural networks have enjoyed successfulimplementation in civil engineering project managementBoussabaine and Elhag [10] developed a fuzzy neural

2 Complexity

network for predicting the duration and cost of constructionworks Yu and Skibniewski (1999) investigated the applica-tion of fuzzy neural networks and genetic algorithms in civilengineering They developed a methodology for the auto-matic collection of experiential data and for detecting factorsthat adversely affect building technology [11] Lam et al [12]successfully applied the principles of fuzzy neural networkstowards creating techniques for modeling uncertainty riskand subjectivity when selecting contractors for the construc-tion works Ko and Cheng [13] developed an evolutionaryfuzzy neural inference model which facilitates decision-making during construction project management Theytested this model using several practical examples duringthe selection of subcontractors for construction works andfor calculating the duration of partition wall constructionan activity that has an excessive impact on the completionof the entire project Jassbi and Khanmohammadi appliedANFIS to risk management [14] Cheng et al proposed animproved hybrid fuzzy neural network for calculating the ini-tial cost of construction works [15] Rashidi et al [16] appliedfuzzy neural systems to the selection of project managers forconstruction projects Mehdi and Reza [17] analyzed theapplication of ANFIS for determining risks in constructionprojects as well as for the development of intelligent systemsfor their assessment Feng and Zhu [18] developed amodel of self-organizing fuzzy neural networks for calculat-ing construction project costs Feylizadeh et al [19] useda model of fuzzy neural networks to calculate completiontime for construction works and to accurately predictdifferent situations

Fuzzy neural networks are also used for an analysis ofstructural elements and structures Ramu and Johnsonapplied the approach of integrating neural networks andfuzzy logic for assessing the damage to composite structures[20] Liu and Wei-guo (2004) investigated the applicationof fuzzy neural networks towards assessing the safety of brid-ges [21] Foncesa used a fuzzy neural system to predict andclassify the behavior of girders loaded with concentratedloads [22] Wang and Liu [23] carried out a risk assessmentin bridge structures using ANFIS Jakubek [24] analyzedthe application of fuzzy neural networks for modeling build-ing materials and the behavior of structures The researchencompassed an analysis of three problems prediction offracture in concrete during fatigue prediction of highperformance concrete strength and prediction of criticalaxial stress for eccentrically loaded reinforced concretecolumns Tarighat [25] developed a fuzzy neural system forassessing risk and damage to bridge structures which allowsimportant information to be predicted related to the impactof design solutions on bridge deterioration Mohammed[26] analyzed the application of fuzzy neural networks inorder to predict the shear strength of ferrocement elementsand concrete girders reinforced with fiber-reinforcedpolymer tapes

Cuumlneyt Aydin et al developed a prognostic model forcalculating the modulus of elasticity for normal damsand high-strength dams with the help of ANFIS [27]Tesfamariam and Najjaran applied the ANFIS model tocalculate the strength of concrete [28] Ozgan et al [29]

developed an adaptive fuzzy neural system (ANFIS) Sugenotype for predicting stiffness parameters for asphalt concrete

Chae and Abraham assessed the state of sewage pipelinesusing a fuzzy neural approach [30] Adeli and Jiang [4]developed a fuzzy neutral model for calculating the capacityof work areas near highways Nayak et al modeled theconnection and the interaction of soil and structures withthe help of fuzzy neural networks Nayak et al [31] appliedANFIS to the hydrological modeling of river flows that isto the forecasting of time-varying data series

Cao and Tian proposed the ANFIS model for predictingthe need for industrial water [32] Chen and Li made a modelfor the quality assessment of river water using the fuzzyneural network methodology [33] F-J Chang and Y-TChang applied a hybrid fuzzy neural approach for construct-ing a system for predicting the water level in reservoirs [34]More precisely they developed a prognostic ANFIS modelfor the management of accumulations while the obtainedresults showed that it could successfully be applied toprecisely and credibly predict the water level in reservoirsJianping et al (2007) developed an improved model of fuzzyneural networks for analysis and deformation monitoring ofdam shifts Hamidian and Seyedpoor have used the method-ology for fuzzy neural networks to determine the optimalshape of arch dams as well as for predicting an effectiveresponse to the impact of earthquakes [35] Thipparat andThaseepetch applied the Sugeno type of ANFIS modelso as to assess structure sustainability of highways inorder to obtain relevant information about environmentalprotection [36]

An increased interest in the application of fuzzy neuralnetworks to civil engineering can be seen in the last fewdecades A comprehensive review of scientific papers whichhave elaborated on this issue published in scientific journalsfrom 1995 until 2017 shows that fuzzy neural networks aremainly used to address several categories of problemsmodeling and predicting calculating and evaluating anddecision-making The results of the conducted analysisillustrate the efficiency and practicality of applying thisinnovative technique towards the development of modelsfor managing decision-making and assessing problemsencountered when planning and implementing construc-tion works Their successful implementation represents apillar for future research within the aforementionedcategories although the future application of fuzzy neuralnetworks can be extended to other areas of civil engineer-ing as well

4 Prognostic Modeling of the Fire Resistance ofEccentrically Loaded Reinforced ConcreteColumns Using Fuzzy Neural Networks

Eccentrically loaded columns are most commonly seen as theend columns in frame structures and they are inserted in thepartition walls that separate the structure from the surround-ing environment or separating the fire compartment underfire conditions The behavior of these types of columns whenexposed to fire and the analysis of the influence of special

3Complexity

factors on their fire resistance have been analyzed inliterature [37 38]

Numerical analysis was carried out for the reinforcedconcrete column (Figure 1) exposed to standard fire ISO834 [27] Due to axial symmetry only one-half of the crosssection was analyzed [37 38] The following input parame-ters were analyzed the dimensions of the cross section theintensity of initial load the thickness of the protectiveconcrete layer the percentage of reinforcement and the typeof concrete (siliceous or carbonate) The output analysisresult is the time period of fire resistance expressed inhours [37 38]

The results from the numerical analysis [37] were used tocreate a prognostic model for determining the fire resistanceof eccentrically loaded reinforced concrete columns in thefire compartment wall

The application of fuzzy neural networks for the determi-nation of fire resistance of eccentrically loaded reinforcedconcrete columns is presented below

The prognostic model was developed using adaptivefuzzy neural networksmdashANFIS in MathWorks softwareusing an integrated Fuzzy Logic Toolbox module [39]

ANFIS represents an adaptive fuzzy neural inferencesystem The advantage of this technique is that membershipfunctions of input parameters are automatically selectedusing a neuroadaptive training technique incorporated intothe Fuzzy Logic Toolbox This technique allows the fuzzymodeling process to learn from the data This is how param-eters of membership functions are calculated through whichthe fuzzy inference system best expresses input-output datagroups [39]

ANFIS represents a fuzzy neural feedforward networkconsisting of neurons and direct links for connectingneurons ANFIS models are generated with knowledge fromdata using algorithms typical of artificial neural networksThe process is presented using fuzzy rules Essentiallyneural networks are structured in several layers throughwhich input data and fuzzy rules are generated Similarto fuzzy logic the final result depends on the given fuzzyrules and membership functions The basic characteristicof ANFIS architecture is that part or all of the neuronsis flexible which means that their output depends on systemparameters and the training rules determine how theseparameters are changed in order to minimize the prescribederror value [40]

ANFIS architecture consists of 5 layers and is illustratedin Figure 2 [3 5ndash7 40]

The first (input) layer of the fuzzy neural network servesto forward input data to the next layer [3 5ndash7 40]

The second layer of the network (the first hidden layer)serves for the fuzzification of input variables Each neuronwithin this layer is represented by the function O1

i = μAi x where x denotes entrance into the neuron i and Ai denoteslinguistic values O1

i is in fact a membership function inAi indicating how many entrances Xi satisfy a quantifierAji The parameters of this layer represent the parameters

of the fuzzy rule premise [3 5ndash7 40]The third layer (the second hidden layer) of the net-

work consists of the T-norm operator for the calculationof fuzzy rule premise Neurons are denoted as π whichrepresents a designation for the product of all input signalswi = μAi x times μBi y Each neuron from this layer establishesthe rule strength of fuzzy rules [3 5ndash7 40]

The fourth network layer (the third hidden layer) nor-malizes the rule strength In each neuron the relationshipbetween the rule strength of the associated rule and thesum of all strengths is calculated wi =wisumwi [3 5ndash7 40]

The procedure for determining subsequent parameters(conclusions) from fuzzy rules is carried out in the fifth layer(the fourth hidden layer) Each node from this layer is asquare (adaptive) node marked by the function O4

i =wiZi =wi piX + qiY + ri where pi qi ri are conclusion parame-ters and wi is the output from the previous layer

The output layer contains one neuron denoted by theletter Σ due to the summing function It calculates the totaloutput as the sum of all input signals in the function ofpremise parameters and fuzzy rule conclusions O5

i =sumwiZi =sumwiZisumwi [3 5ndash7 40]

For a fuzzy neural network consisting of 2 input variablesand 2 fuzzy rules (Figure 2) the total output would becalculated as follows [3 5ndash7 40]

Z =wiZi =sumwiZi

sumwi= w1w1 +w2

Z1 +w2

w1 +w2Z1

=w1Z1 +w2Z2 =w1 p1X + q1Y + r1+w2 p2X + q2Y + r2

1

where X Y is the numerical input of the fuzzy neural net-work Z is the numerical output of the fuzzy neural networkw1w2 are the normalized rule strengths of fuzzy rulesexpressed through the fuzzy rule premise and p1 q1 r1p2 q2 r2 are the parameters of the fuzzy rule conclusions

The ANFIS training algorithm consists of two segmentsreverse propagation method (backpropagation algorithm)which determines errors of variables from a recursive pathfrom the output to the input layers determining variableerrors that is parameters of the membership function andthe least square method determining the optimal set ofconsequent parameters Each step in the training procedureconsists of two parts In the first part input data is propa-gated and optimal consequent parameters are estimatedusing the iterative least-mean-square method while fuzzyrule premise parameters are assumed to be fixed for thecurrent cycle through the training set In the second partinput data is propagated again but in this process the

N

M

Bel 3

el 2

el 1A

h = 3 m

b

d2

d2

y Tf = 20degC

Figure 1 RC column inserted into the fire separation wall

4 Complexity

backpropagation algorithm is used to modify the premiseparameter while the consequent parameters remain fixedThis procedure is iterated [3 5ndash7 40]

For the successful application of ANFIS during theprocess of solving a specific problem task it is necessary topossess solid professional knowledge of the problem at handand appropriate experience This enables a correct andaccurate choice of input variables that is unnecessarilycomplicating the model by adding nonsignificant variablesor not including important parameters that have a significanteffect on output values is avoided

The application of fuzzy neural network techniques tothe modeling process is carried out in a few steps [39 41]assembling and processing data determining the parametersand structure of the fuzzy neural network (creating the fuzzyinference system) training the fuzzy neural network andtesting the fuzzy neural network and prognostics

For the purpose of prognostic modeling of the eccentri-cally loaded RC columns the structure of the fuzzy neuralnetwork consists of 6 input variables (dimensions of thereinforced concrete column (b and d) thickness of the pro-tective concrete layer (a) percentage of reinforcement (μ)axial load coefficient (η) bending moment coefficient (β)and one output variable (fire resistance of the reinforcedconcrete column (t))

One of the most crucial aspects when using a fuzzyneural networks as prognostic modeling technique is tocollect accurate and appropriate data sets The data hasto contain a finite number of sets where each data sethas to be defined with an exact input and output valuesAnother very important aspect is to have large amountof data sets Data sets are divided into two main groupsdata used for training of the model and data used for testingof the model prediction accuracy The training data shouldcontain all the necessary representative features so the pro-cess of selecting a data set for checking or testing purposesis made easier One problem for models constructed usingadaptive techniques is selecting a data set that is both repre-sentative of the data the trained model is intended to imitateyet sufficiently distinct from the training data set so as not to

render the validation process trivial To design an ANFISsystem for real-world problems it is essential to select theparameters for the training process It is essential to haveproper training and testing data sets If the data sets are notselected properly then the testing data set will not validatethe model If the testing data set is completely differentfrom the training data set then the model cannot captureany of the features of the testing data Then the minimumtesting error can be achieved in the first epoch For the properdata set the testing error decreases with the training proceed-ing until a jump point The selection of data sets for trainingand testing of the ANFIS system is an important factor affect-ing the performance of the model If the data sets used fortesting is extremely different from one of the training datasets then the system fails to capture the essential features ofthe data set Another aspect that has to be emphasized is thatall data sets have to be properly selected adequately collectedand exact The basic characteristic of all computer programsused for calculation and modeling applies for the neuralnetworks as well only quality input can give a quality outputEven though neural networks represent an intelligentmodeling technique they are not omnipotent which meansthat if the input data sets are not clear and correct theneural network model will not be able to produce accurateoutput results

A training data group is used to initially create a modelstructure [39] The training process is a learning process ofthe developed model The model is trained till the resultsare obtained with minimum error During the learningprocess the parameters of the membership functions areupdated In MATLAB the two ANFIS parameter optimiza-tion methods are hybrid (combination of least squares andback propagation method) and back propagation Errortolerance is used as training stopping criterion which isrelated to the error size The training will stop after the train-ing data error remains within this tolerance The trainingerror is the difference between the training data output valueand the output of the fuzzy inference system correspondingto the same training data input value (the one associated withthat training data output value)

Inputlayer

Hiddenlayer 1

Hiddenlayer 2

Hiddenlayer 3

Hiddenlayer 4

Outputlayer

X

Y

Premiseparameters

Conclusionparameters

A1

w1

w2

w1

w2

w2Z2

w1Z1

x1 x2

x1 x2

A2

B1

B2

N

N

120587

120587

f

f

sumZ

Figure 2 An overview of the ANFIS network

5Complexity

The data groups used for checking and testing of themodel are also referred as validation data and are used tocheck the capabilities of the generalization model duringtraining [39] Model validation is the process by which theinput data sets on which the FIS was not trained are pre-sented to the trained FIS model to see the performanceThe validation process for the ANFIS model is carried outby leaking vectors from the input-output testing data intothe model (data not belonging to the training group) in orderto verify the accuracy of the predicted output Testing data isused to validate the model The testing data set is used forchecking the generalization capability of the ANFIS modelHowever it is desirable to extract another group of input-output data that is checking data in order to avoid thepossibility of an ldquooverfittingrdquo The idea behind using thechecking data stems from the fact that a fuzzy neural networkmay after a certain number of training cycles be ldquoover-trainedrdquo meaning that the model practically copies theoutput data instead of anticipating it providing great predic-tions but only for training data At that point the predictionerror for checking data begins to increase when it shouldexhibit a downward trend A trained network is tested withchecking data and parameters for membership functionsare chosen for which minimal errors are obtained These datacontrols this phenomenon by adjusting ANFIS parameterswith the aim of achieving the least errors in predictionHowever when selecting data for model validation a detaileddatabase study is required because the validation data shouldbe not only sufficiently representative of the training databut also sufficiently different to avoid marginalization oftraining [39 41]

Even though there are many published researches world-wide that investigate the impact of the proportion of dataused in various subsets on neural network model there isno clear relationship between the proportion of data fortraining testing and validation and model performanceHowever many authors recommend that the best result canbe obtained when 20 of the data are used for validationand the remaining data are divided into 70 for trainingand 30 for testing The number of training and testing datasets greatly depends on the total number of data sets So thereal apportion of train and test data set is closely related with

the real situation problem specifics and the quantity of thedata set There are no strict rules regarding the data setdivision so when using the adaptive modeling techniquesit is very important to know how well the data sets describethe features of the problem and to have a decent amount ofexperience and knowledge about neural networks [42 43]

The database used for ANFIS modeling for the modelpresented in this paper expressed through input and outputvariables was obtained from a numerical analysis A total of398 input-output data series were analyzed out of which318 series (80) were used for network training and 80 series(20) were data for testing the model The prediction of theoutput result was performed on new 27 data sets The threedata groups loaded into ANFIS are presented in Figures 3ndash5

For a precise and reliable prediction of fire resistance foreccentrically loaded reinforced concrete columns differentANFIS models were analyzed with the application of thesubtractive clustering method and a hybrid training mode

The process of training fuzzy neural networks involvesadjusting the parameters of membership functions Trainingis an iterative process that is carried out on training data setsfor fuzzy neural networks [39] The training process endswhen one of the two defined criteria has been satisfied theseare error tolerance percentage and number of iterations Forthe analysis in this research the values of these criteria were 0for error tolerance and 100 for number of training iterations

After a completed training process of the generatedANFIS model it is necessary to validate the model using datasets defined for testing and checking [39 44] the trainedmodel is tested using validation data and the obtainedaverage errors and the estimated output values are analyzed

The final phase of the modeling using fuzzy neuralnetworks is prognosis of outputs and checking the modelrsquosprediction accuracy To this end input data is passed throughthe network to generate output results [39 44] If low valuesof average errors have been obtained during the testingand validation process of the fuzzy neural network thenit is quite certain that a trained and validated ANFISmodel can be applied for a high-quality and precise prog-nosis of output values

For the analyzed case presented in this paper the optimalANFIS model is determined by analyzing various fuzzy

Training data (ooo)25

20

15

10Out

put

5

00 50 100 150 200

Data set index250 300 350

Figure 3 Graphical representation of the training data loaded into ANFIS

6 Complexity

neural network architectures obtained by varying the follow-ing parameters the radius of influence (with values from 06to 18) and the compaction factor (with values in the intervalfrom 045 to 25) The mutual acceptance ratio and themutual rejection ratio had fixed values based on standardvalues defined in MATLAB amounting to 05 and 015[39] A specific model of the fuzzy neural network witha designation depending on the parameter values wascreated for each combination of these parameters Theoptimal combination of parameters was determined byanalyzing the behavior of prognostic models and by com-paring obtained values for average errors during testingand predicting output values Through an analysis of theresults it can be concluded that the smallest value foraverage error in predicting fire resistance of eccentricallyloaded reinforced concrete columns is obtained using theFIS11110 model with Gaussian membership function(gaussMF) for the input variable and linear output functionwith 8 fuzzy rules The average error during model testingwas 0319 for training data 0511 for testing data and 0245for verification data

Figure 6 gives a graphical representation of the architec-ture of the adopted ANFIS model composed of 6 input

Checking data (+++)12

10

8

Out

put

4

6

2

00 5 10 15 20

Data set index25 30

Figure 5 Graphical representation of the checking data loaded into ANFIS

Testing data ()20

15

10

Out

put

5

00 10 20 30 40

Data set index50 60 70 80

Figure 4 Graphical representation of the test data loaded into ANFIS

Input Inputmf Outputmf OutputRule

Logical operationsAndOrNot

Figure 6 Architecture of ANFIS model FIS11110 composed of6 input variables and 1 output variable

7Complexity

variables defined by Gaussian membership functions and1 output variable

The training of the ANFIS model was carried out with318 input-output data groups A graphic representation offire resistance for the analyzed reinforced concrete columnsobtained by numerical analysis [37] and the predicted valuesobtained by the FIS11110 prognostic model for the trainingdata is given in Figure 7

It can be concluded that the trained ANFIS prognosticmodel provides excellent results and quite accurately predictsthe time of fire resistance of the analyzed reinforced concretefor input data that belong to the size intervals the networkwas trained formdashthe analyzed 318 training cases The averageerror that occurs when testing the network using trainingdata is 0319

The testing of the ANFIS model was carried out using 80inputoutput data sets A graphical comparison representa-tion of the actual and predicted values of fire resistance foranalyzed reinforced concrete columns for the testing datasets is presented in Figure 8

Figures 7 and 8 show an excellent match between pre-dicted values obtained from the ANFIS model with the actualvalues obtained by numerical analysis [37] The average error

that occurs when testing the fuzzy neural network using test-ing data is 0511

Figure 9 presents the fuzzy rules as part of the FuzzyLogic Toolbox program of the trained ANFIS modelFIS11110 Each line in the figure corresponds to one fuzzyrule and the input and output variables are subordinated inthe columns Entering new values for input variables auto-matically generates output values which very simply predictsfire resistance for the analyzed reinforced concrete columns

The precision of the ANFIS model was verified using 27input-output data groups (checking data) which also repre-sents a prognosis of the fuzzy neural network because theywere not used during the network training and testing Theobtained predicted values for fire resistance for the analyzedreinforced concrete columns are presented in Table 1 Agraphic representation of the comparison between actualvalues (obtained by numerical analysis [37]) and predictedvalues (obtained through the ANFIS model FIS11110) for fireresistance of eccentrically loaded reinforced concrete col-umns in the fire compartment wall is presented in Figure 10

An analysis of the value of fire resistance for the analyzedeccentrically loaded reinforced concrete columns shows thatthe prognostic model made with fuzzy neural networks

Training data o FIS output 25

20

15

Out

put

5

10

00 50 100 150 200 250

Index300 350

Figure 7 Comparison of the actual and predicted values of fire resistance time for RC columns obtained with ANFIS training data

Testing data 20

15

Out

put

0

5

10

minus50 10 20 30 40 50 60

Index70 80

FIS output

Figure 8 A comparison of actual and predicted values of fire resistance time for RC columns obtained from the ANFIS testing data

8 Complexity

provides a precise and accurate prediction of output resultsThe average square error obtained when predicting outputresults using a fuzzy neural network (ANFIS) is 0242

This research indicates that this prognostic modelenables easy and simple determination of fire resistance ofeccentrically loaded reinforced concrete columns in the firecompartment wall with any dimensions and characteristics

Based on a comparison of results obtained from a numer-ical analysis and results obtained from the prognostic modelmade from fuzzy neural networks it can be concluded thatfuzzy neural networks represent an excellent tool for deter-mining (predicting) fire resistance of analyzed columnsThe prognostic model is particularly useful when analyzingcolumns for which there is no (or insufficient) previousexperimental andor numerically derived data and a quickestimate of its fire resistance is needed A trained fuzzy neuralnetwork gives high-quality and precise results for the inputdata not included in the training process which means thata projected prognostic model can be used to estimate rein-forced concrete columns of any dimension and characteristic(in case of centric load) It is precisely this positive fact thatfully justifies the implementation of more detailed andextensive research into the application of fuzzy neural net-works for the design of prognostic models that could be usedto estimate different parameters in the construction industry

5 Conclusion

Prognostic models based on the connection between popularmethods for soft computing such as fuzzy neural networksuse positive characteristics of neural networks and fuzzysystems Unlike traditional prognostic models that work

precisely definitely and clearly fuzzy neural models arecapable of using tolerance for inaccuracy uncertainty andambiguity The success of the ANFIS is given by aspects likethe designated distributive inferences stored in the rule basethe effective learning algorithm for adapting the systemrsquosparameters or by the own learning ability to fit an irregularor nonperiodic time series The ANFIS is a technique thatembeds the fuzzy inference system into the framework ofadaptive networks The ANFIS thus draws the benefits ofboth ANN and fuzzy techniques in a single frameworkOne of the major advantages of the ANFIS method overfuzzy systems is that it eliminates the basic problem ofdefining the membership function parameters and obtaininga set of fuzzy if-then rules The learning capability of ANN isused for automatic fuzzy if-then rule generation and param-eter optimization in the ANFIS The primary advantages ofthe ANFIS are the nonlinearity and structured knowledgerepresentation Research and applications on fuzzy neuralnetworks made clear that neural and fuzzy hybrid systemsare beneficial in fields such as the applicability of existingalgorithms for artificial neural networks (ANNs) and directadaptation of knowledge articulated as a set of fuzzy linguis-tic rules A hybrid intelligent system is one of the bestsolutions in data modeling where it is capable of reasoningand learning in an uncertain and imprecise environment Itis a combination of two or more intelligent technologies Thiscombination is done usually to overcome single intelligenttechnologies Since ANFIS combines the advantages of bothneural network and fuzzy logic it is capable of handlingcomplex and nonlinear problems

The application of fuzzy neural networks as an uncon-ventional approach for prediction of the fire resistance of

in1 = 40 in2 = 40 in3 = 3 in4 = 105 in5 = 03 in6 = 025out1 = 378

1580minus2006

1

2

3

4

5

6

7

8

30

Input[4040310503025]

Plot points101

Moveleft

Help Close

right down up

Opened system Untitled 8 rules

50 30 50 2 4 06 15 01 05 050

Figure 9 Illustration of fuzzy rules for the ANFIS model FIS11110

9Complexity

Checking data + FIS output 25

20

15

Out

put

5

10

00 5 10 15 20

Index25 30

Figure 10 Comparison of actual and predicted values of fire resistance time of RC columns obtained using the ANFIS checking data

Table 1 Actual and predicted value of fire resistance time for RC columns obtained with the ANFIS checking data

Checking data

Columndimensions

Thickness of theprotective concrete layer

Percentage ofreinforcement

Axial loadcoefficient

Bending momentcoefficient

Fire resistance time of eccentricallyloaded RC columns

Actual values Predicted values (ANFIS)b d a μ η β t t

3000 3000 200 100 010 02 588 553

3000 3000 200 100 030 03 214 187

3000 3000 300 100 010 05 294 344

3000 3000 300 100 040 04 132 132

3000 3000 400 100 040 01 208 183

4000 4000 200 100 010 02 776 798

4000 4000 200 100 040 0 382 383

4000 4000 300 100 020 04 356 374

4000 4000 300 100 030 05 216 228

4000 4000 300 100 040 0 38 377

4000 4000 400 100 010 01 1014 99

4000 4000 400 100 030 03 344 361

4000 4000 300 060 020 02 556 552

4000 4000 300 150 010 0 12 117

4000 4000 300 150 050 04 138 128

5000 5000 200 100 010 03 983 101

5000 5000 300 100 010 05 528 559

5000 5000 400 100 030 04 384 361

5000 5000 200 060 020 01 95 903

5000 5000 200 060 050 0 318 337

5000 5000 400 060 030 02 57 553

5000 5000 400 060 050 0 352 353

5000 5000 400 060 050 01 314 313

5000 5000 400 060 050 02 28 281

5000 5000 400 060 050 03 232 252

5000 5000 400 060 050 04 188 22

5000 5000 400 060 050 05 148 186

10 Complexity

structural elements has a huge significance in the moderniza-tion of the construction design processes Most of the exper-imental models for the determination of fire resistance areextremely expensive and analytical models are quite compli-cated and time-consuming That is why a modern type ofanalyses such as modeling through fuzzy neural networkscan help especially in those cases where some prior analyseswere already made

This paper presents some of the positive aspects of theirapplication for the determination the fire resistance ofeccentrically loaded RC columns exposed to standard firefrom one side The influence of the cross-sectional dimen-sions thickness of the protective concrete layer percentageof reinforcement and the intensity of the applied loads tothe fire resistance of eccentrically loaded RC columns wereanalyzed using the program FIRE The results of the per-formed numerical analyses were used as input parametersfor training of the ANFIS model The obtained outputsdemonstrate that the ANFIS model is capable of predictingthe fire resistance of the analyzed RC columns

The results from this research are proof of the successfulapplication of fuzzy neural networks for easily and simplysolving actual complex problems in the field of constructionThe obtained results as well as the aforementioned conclud-ing considerations emphasize the efficiency and practicalityof applying this innovative technique for the developmentof management models decision making and assessmentof problems encountered during the planning and imple-mentation of construction projectsworks

A fundamental approach based on the application offuzzy neural networks enables advanced and successfulmodeling of fire resistance of reinforced concrete columnsembedded in the fire compartment wall exposed to fire onone side thus overcoming defects typical for traditionalmethods of mathematical modeling

Data Availability

The data used to support the findings of this study areavailable from the corresponding author upon request


The authors declare that they have no conflicts of interest

References

[1] EN1994ndash1-2 Eurocode 4 ndash Design of Composite Steel andConcrete Structures Part 1-2 General Rules - Structural FireDesign European Committee for Standardization Manage-ment Centre 2005

[2] A Abraham and B Nath ldquoA neuro-fuzzy approach for model-ling electricity demand in Victoriardquo Applied Soft Computingvol 1 no 2 pp 127ndash138 2001

[3] A Abrahim ldquoNeuro fuzzy systems state-of-the-art modelingtechniquesrdquo in Connectionist Models of Neurons LearningProcesses and Artificial Intelligence Lecture notes in Computersciences J Mira and A Prieto Eds pp 269ndash276 Springer-Verlag Germany 2001

[4] H Adeli and X Jiang ldquoNeuro-fuzzy logic model for freewaywork zone capacity estimationrdquo Journal of TransportationEngineering vol 129 no 5 pp 484ndash493 2003

[5] A Abraham It Is Time to Fuzzify Neural Networks (Tutorial)International Conference on Intelligent Multimedia andDistance Education Fargo USA 2001

[6] D Fuller Neural Fuzzy Systems Abo Akademi University1995

[7] M F Azeem Ed Fuzzy Inference System-Theory andApplications InTechOpen 2012

[8] N K Kasabov ldquoFoundations of neural networks fuzzysystems and knowledge engineeringrdquo in A Bradford BookThe MIT Press Cambridge London England 1998

[9] A Abrahim and M R Khan ldquoNeuro-fuzzy paradigms forintelligent energy managementrdquo in Connectionist Models ofNeurons Learning Processes and Artificial Intelligence Lecturenotes in Computer sciences J Mira and A Prieto EdsSpringer-Verlag Berlin Heidelberg 2004

[10] A H Boussabaine and T M S Elhag A Neurofuzzy Model forPredicting Cost and Duration of Construction Projects RoyalInstitution of Chartered Surveyors 1997

[11] S S H Yasrebi andM Emami ldquoApplication of artificial neuralnetworks (ANNs) in prediction and interpretation of Pres-suremeter test resultsrdquo in The 12th International Conferenceof International Association for Computer Methods andAdvances in Geomechanics (IACMAG) pp 1634ndash1638 India2008

[12] K C Lam T Hu S Thomas Ng M Skitmore and S OCheung ldquoA fuzzy neural network approach for contractorprequalificationrdquo Construction Management and Economicsvol 19 no 2 pp 175ndash188 2001

[13] C H Ko and M Y Cheng ldquoHybrid use of AI techniques indeveloping construction management toolsrdquo Automation inConstruction vol 12 no 3 pp 271ndash281 2003

[14] J Jassbi and S Khanmohammadi Organizational RiskAssessment Using Adaptive Neuro-Fuzzy Inference SystemIFSA-EUSFLAT 2009

[15] M-Y Cheng H-C Tsai C-H Ko and W-T ChangldquoEvolutionary fuzzy neural inference system for decisionmaking in geotechnical engineeringrdquo Journal of Computingin Civil Engineering vol 22 no 4 pp 272ndash280 2008

[16] A Rashidi F Jazebi and I Brilakis ldquoNeurofuzzy geneticsystem for selection of construction project managersrdquo Journalof Construction Engineering and Management vol 137 no 1pp 17ndash29 2011

[17] E Mehdi and G Reza ldquoRisk assessment of constructionprojects using network based adaptive fuzzy systemrdquo Inter-national Journal of Academic Research vol 3 no 1 p 4112011

[18] W F Feng and W J Zhu ldquoThe application of SOFM fuzzyneural network in project cost estimaterdquo Journal of Softwarevol 6 no 8 pp 1452ndash1459 2011

[19] M R Feylizadeh A Hendalianpour and M BagherpourldquoA fuzzy neural network to estimate at completion costsof construction projectsrdquo International Journal of Indus-trial Engineering Computations vol 3 no 3 pp 477ndash484 2012

[20] S A Ramu and V T Johnson ldquoDamage assessment ofcomposite structuresmdasha fuzzy logic integrated neural networkapproachrdquo Computers amp Structures vol 57 no 3 pp 491ndash502 1995

11Complexity

[21] M Y Liu and Y Wei-guo ldquoResearch on safety assessment oflong-span concrete-filled steel tube arch bridge based onfuzzy-neural networkrdquo China Journal of Highway andTransport vol 4 p 012 2004

[22] E T Foncesa A Neuro-Fuzzy System for Steel Beams PatchLoad Prediction Fifth International Conference on HybridIntelligent Systems 2005

[23] B Wang X Liu and C Luo Research on the Safety Assessmentof Bridges Based on Fuzzy-Neural Network Proceedings ofthe Second International Symposium on Networking andNetwork Security China 2010

[24] M Jakubek ldquoFuzzy weight neural network in the analysisof concrete specimens and RC column buckling testsrdquoComputer Assisted Methods in Engineering and Sciencevol 18 no 4 pp 243ndash254 2017

[25] A Tarighat ldquoFuzzy inference system as a tool for managementof concrete bridgesrdquo in Fuzzy Inference System-Theory andApplications M F Azeem Ed IntechOpen 2012

[26] M A Mashrei ldquoNeural network and adaptive neuro-fuzzyinference system applied to civil engineering problemsrdquo inFuzzy Inference System-Theory and Applications IntechOpen2012

[27] A Cuumlneyt Aydin A Tortum and M Yavuz ldquoPrediction ofconcrete elastic modulus using adaptive neuro-fuzzy inferencesystemrdquo Civil Engineering and Environmental Systems vol 23no 4 pp 295ndash309 2006

[28] S Tesfamariam and H Najjaran ldquoAdaptive networkndashfuzzyinferencing to estimate concrete strength using mix designrdquoJournal of Materials in Civil Engineering vol 19 no 7pp 550ndash560 2007

[29] E Oumlzgan İ Korkmaz and M Emiroğlu ldquoAdaptive neuro-fuzzy inference approach for prediction the stiffness moduluson asphalt concreterdquo Advances in Engineering Softwarevol 45 no 1 pp 100ndash104 2012

[30] M J Chae and D M Abraham ldquoNeuro-fuzzy approaches forsanitary sewer pipeline condition assessmentrdquo Journal ofComputing in Civil Engineering vol 15 no 1 pp 4ndash14 2001

[31] P C Nayak K P Sudheer DM Rangan and K S RamasastrildquoA neuro-fuzzy computing technique for modelinghydrological time seriesrdquo Journal of Hydrology vol 291no 1-2 pp 52ndash66 2004

[32] A Cao and L Tian ldquoApplication of the adaptive fuzzy neuralnetwork to industrial water consumption predictionrdquo Journalof Anhui University of Technology and Science vol 3 2005

[33] S Y Chen and Y W Li ldquoWater quality evaluation based onfuzzy artificial neural networkrdquo Advances in Water Sciencevol 16 no 1 pp 88ndash91 2005

[34] F-J Chang and Y-T Chang ldquoAdaptive neuro-fuzzy inferencesystem for prediction of water level in reservoirrdquo Advances inWater Resources vol 29 no 1 pp 1ndash10 2006

[35] D Hamidian and S M Seyedpoor ldquoShape optimal designof arch dams using an adaptive neuro-fuzzy inferencesystem and improved particle swarm optimizationrdquo AppliedMathematical Modelling vol 34 no 6 pp 1574ndash1585 2010

[36] T Thipparat and T Thaseepetch ldquoApplication of neuro-fuzzysystem to evaluate sustainability in highway desighrdquo Interna-tional Journal of Modern Engineering Research vol 2 no 5pp 4153ndash4158 2012

[37] M Cvetkovska Nonlinear Stress Strain Behavior of RCElements and Plane Frame Structures Exposed to Fire Doctoral

Dissertation Civil Engineering Faculty in Skopje ldquoSts Cyriland Methodiusrdquo University Macedonia 2002

[38] M Lazarevska M Knežević M Cvetkovska N IvaniševićT Samardzioska and A T Gavriloska ldquoFire-resistance prog-nostic model for reinforced concrete columnsrdquo Građevinarvol 64 p 7 2012

[39] httpwwwmathworkscomproductsfuzzy-logic

[40] J-S R Jang ldquoANFIS adaptive-network-based fuzzy inferencesystemrdquo IEEE Transactions on Systems Man and Cyberneticsvol 23 no 3 pp 665ndash685 1993

[41] J Guan J Zurada and S Levitian ldquoAn adaptive neuro-fuzzyinference system based approach to real estate propertyassessmentrdquo Journal of Real Estate Research vol 30 no 4pp 395ndash421 2008

[42] H Adeli ldquoNeural networks in civil engineering 1989ndash2000rdquoComputer-Aided Civil and Infrastructure Engineering vol 16no 2 pp 126ndash142 2001

[43] E B Baum and D Haussler ldquoWhat size net gives valid gener-alizationrdquo Neural Computation vol 1 no 1 pp 151ndash1601989

[44] M Neshat A Adela A Masoumi and M Sargolzae ldquoAcomparative study on ANFIS and fuzzy expert system modelsfor concrete mix designrdquo International Journal of ComputerScience Issues vol 8 no 2 pp 196ndash210 2011

12 Complexity

Research ArticleUrban Road Infrastructure Maintenance Planning withApplication of Neural Networks

Ivan Marović 1 Ivica Androjić 1 Nikša Jajac2 and Tomaacuteš Hanaacutek 3

1Faculty of Civil Engineering University of Rijeka Radmile Matejčić 3 51000 Rijeka Croatia2Faculty of Civil Engineering Architecture and Geodesy University of Split Matice hrvatske 15 21000 Split Croatia3Faculty of Civil Engineering Brno University of Technology Veveri 95 602 00 Brno Czech Republic

Correspondence should be addressed to Tomaacuteš Hanaacutek hanaktfcevutbrcz

Received 23 February 2018 Revised 11 April 2018 Accepted 12 April 2018 Published 29 May 2018

Academic Editor Lucia Valentina Gambuzza

Copyright copy 2018 Ivan Marović et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

The maintenance planning within the urban road infrastructure management is a complex problem from both the managementand technoeconomic aspects The focus of this research is on decision-making processes related to the planning phase duringmanagement of urban road infrastructure projects The goal of this research is to design and develop an ANN model in order toachieve a successful prediction of road deterioration as a tool for maintenance planning activities Such a model is part of theproposed decision support concept for urban road infrastructure management and a decision support tool in planning activitiesThe input data were obtained from Circly 60 Pavement Design Software and used to determine the stress values (560 testingcombinations) It was found that it is possible and desirable to apply such a model in the decision support concept in order toimprove urban road infrastructure maintenance planning processes

1 Introduction

The development of urban road infrastructure systems is anintegral part of modern city expansion processes Interna-tionally roads are dominant transport assets and a valuableinfrastructure used on a daily basis by millions of commuterscomprising millions of kilometers across the world Accord-ing to [1] the average length of public roads in OECD coun-tries is more than 500000 km and is often the single largestpublicly owned national asset Such infrastructure covers15ndash20 of the whole city area and in city centers over 40of the area [2] Therefore the road infrastructure is unargu-ably seen as significant and valuable public asset whichshould be carefully managed during its life cycle

In general the importance of road maintenance can beseen as the following [1]

(i) Roads are key national assets which underpin eco-nomic activity

(ii) Road transport is a foundation for economicactivity

(iii) Ageing infrastructure requires increased roadmaintenance

(iv) Traffic volumes continue to grow and driveincreased need for maintenance

(v) Impacts of road maintenance are diverse and mustbe understood

(vi) Investing in maintenance at the right time savessignificant future costs

(vii) Maintenance investment must be properlymanaged

(viii) Road infrastructure planning is imperative for roadmaintenance for future generations

In urban areas the quality of road infrastructure directlyinfluences the citizensrsquo quality of life [3] such as the resi-dentsrsquo health safety economic opportunities and conditionsfor work and leisure [3 4] Therefore every action needscareful planning as it is highly complex and socially sensitiveIn order to deal with such problems city governments often


encounter considerable problems during the planning phasewhen it is necessary to find a solution that would meet therequirements of all stakeholders and at the same time be apart of the desired development concept As they are limitedby certain annual budgeting for construction maintenanceand remedial activities the projectrsquos prioritization emergesas one of the most important and most difficult issues to beresolved in the public decision-making process [5]

In order to cope with such complexity various manage-ment information systems were created Some aimed atimproving decision-making at the road infrastructure plan-ning level in urban areas based on multicriteria methods(such as simple additive weighting (SAW) and analytichierarchy processing (AHP)) and artificial neural networks(ANNs) [5] others on combining several multicriteriamethods (such as AHP and PROMETHEE [6] AHP ELEC-TRE and PROMETHEE [7]) or just using single multicri-teria method (such as AHP [8]) Deluka-Tibljaš et al [2]reviewed various multicriteria analysis methods and theirapplication in decision-making processes regarding trans-port infrastructure They concluded that due to complexityof the problem application of multicriteria analysis methodsin systems such as decision support system (DSS) can signif-icantly contribute to the improvement of the quality ofdecision-making process regarding transport infrastructurein urban areas

Apart from the aforementioned systems which aremainly used for strategic management most maintenancemanagement aspects are connected to various pavement sys-tems A typical pavement management system should help adecision-maker to select the best maintenance program sothat the maximal use is made of available resources Such aprogram answers questions such as which maintenancetreatment to use and where and when to apply it The qualityof the prioritization directly influences the effectiveness ofavailable resources which is often the primary decision-makersrsquo goal Therefore Wang et al [9] developed an integerlinear programming model in order to select a set of candi-date projects from the highway network over a planninghorizon of 5 years Proposed model was tested on a smallnetwork of 10 road sections regarding two optimizationobjectivesmdashmaximization of the total maintenance andrehabilitation effectiveness and minimization of the totalmaintenance and rehabilitation disturbance cost For yearspavement management systems have been used in highwayagencies to improve the planning efforts associated withpavement preservation activities to provide the informationneeded to support the pavement preservation decision pro-cess and to compare the long-term impacts of alternativepreservation strategies As such pavement management isan integral part of an agencyrsquos asset management effortsand an important tool for cost-effectively managing the largeinvestment in its transportation infrastructure Zimmermanand Peshkin [10] emphasized the issues regarding integratingpavement management and preventive maintenance withrecommendations for improving pavement management sys-tems while Zhang et al [11] developed a new network-levelpavement asset management system utilizing life cycle analy-sis and optimization methods The proposed management

systemallowsdecision-makers topreserve ahealthypavementnetwork and minimize life cycle energy consumption green-house gas emission or cost as a single objective and also meetbudget constraints and other decision-makerrsquos constraints

Pavements heavily influence the management costs inroad networks Operating pavements represent a challengingtask involving complex decisions on the application of main-tenance actions to keep them at a reasonable level of perfor-mance The major difficulty in applying computational toolsto support decision-making lies in a large number of pave-ment sections as a result of a long length of road networksTherefore Denysiuk et al [12] proposed a two-stage multi-objective optimization of maintenance scheduling for pave-ments in order to obtain a computationally treatable modelfor large road networks As the given framework is generalit can be extended to different types of infrastructure assetsAbo-Hashema and Sharaf [13] proposed a maintenance deci-sion model for flexible pavements which can assist decision-makers in the planning and cost allocation of maintenanceand rehabilitation processes more effectively They developa maintenance decision model for flexible pavements usingdata extracted from the long-term pavement performanceDataPave30 software The proposed prediction model deter-mines maintenance and rehabilitation activities based on thedensity of distress repair methods and predicts future main-tenance unit values with which future maintenance needsare determined

Application of artificial neural networks in order todevelop prediction models is mostly connected to road mate-rials and modelling pavement mixtures [14ndash16] rather thanplanning processes especially maintenance planning There-fore the goal of this research is to design and develop anANN model in order to achieve a successful prediction ofroad deterioration as a tool for maintenance planning activi-ties Such a model is part of the proposed decision supportconcept (DSC) for urban road infrastructure managementand a decision support tool in planning activities

This paper is organized as follows Section 2 provides aresearch background of the decision support concept as wellas the methodology for the development of ANN predictionmodels as a tool for supporting decisions in DSC In Section3 the results of the proposed model are shown and discussedFinally the conclusion and recommendations are presentedin Section 4

2 Methodology

21 Research Background Depending on the need of thebusiness different kinds of information systems are devel-oped for different purposes Many authors have studied pos-sibilities for generating decision support tools for urbanmanagement in the form of various decision support sys-tems Such an approach was done by Bielli [17] in order toachieve maximum efficiency and productivity for the entireurban traffic system while Quintero et al [18] described animproved version of such a system named IDSS (intelligentdecision support system) as it coordinates management ofseveral urban infrastructure systems at the same time Jajacet al [5 6] presented how different decision support models

2 Complexity

can be generated and used at different decision-making levelsfor the purpose of cost-and-benefit analyses of potentialinfrastructure investments A decision support concept(DSC) aimed at improving urban road infrastructure plan-ning based on multicriteria methods and artificial neural net-works proposed by Jajac et al [4] showed how urban roadinfrastructure planning can be improved It showed howdecision-making processes during the planning stages canbe supported at all decision-making levels by proper interac-tion between DSC modules

A structure of the proposed decision support concept forurban road infrastructure management (Figure 1) is based onthe authorrsquos previous research where the ldquothree decisionlevelsrdquo concept for urban infrastructure [5 6 19] and spatial[4 20] management is proposed The proposed concept ismodular and based on DSS basic structure [21] (i) database(ii) model base and (iii) dialog module Interactions betweenmodules are realized throughout the decision-making pro-cess at all management levels as they serve as meeting pointsof adequate models from the model base and data from thedatabase module Also interactions between decision-makers experts and stakeholders are of crucial importanceas they deal with various types of problems (from structuredto unstructured)

The first management level supports decision-makers atthe lowest operational management level Besides its generalfunction of supporting decision-making processes at theoperational level it is a meeting point of data and informa-tion where the problems are well defined and structuredAdditionally it provides information flows towards higherdecision levels (arrow 1 in Figure 1) The decision-makersat the second management level (ie tactical managementlevel) deal with less-defined and semistructured problemsAt this level tactical decisions are delivered and it is a placewhere information basis and solutions are created Based onapplied models from the model base it gives alternatives anda basis for future decisions on the strategic management level(arrow 2 in Figure 1) which deals with even less-defined andunstructured problems Depending on the decision problemvarious methods could be used (ANN eg) At the thirdmanagement level based on the expert deliverables fromthe tactical level a future development of the system is car-ried out Strategies are formed and they serve as frameworksfor lower decision and management levels (arrows 3 and 4 inFigure 1)

As the decisions made throughout the system are basedon knowledge generated at the operational decision-makinglevel it is structured in an adequate knowledge-based systemof the database (geographic information system (GIS) eg)Besides DSS structure elements which obviously influencethe system at all management levels other factors from theenvironment have a considerable influence on both thedecision-making process as well as management processesas is shown in Figure 1 Such structure is found to be ade-quate for various urban management systems and its struc-ture easily supports all phases of the decision-makingprocess Since this research is focused on the urban roadinfrastructure management and particularly on the improve-ment of its planning process the concept is used to support

decision-making processes in the realization of these man-agement functions In this paper the focus is on the mainte-nance planning process of the urban road infrastructure withthe application of neural networks

The development of the ANN prediction model requiresa number of technologies and areas of expertise It is neces-sary to collect an adequate set of data (from monitoringandor collection of existing historical data) from the urbanarea From such a collection the data analysis is made whichserves as a starting point on the prediction model develop-ment Importing such a prediction model into an existingor the development of a new decision support system resultsin a supporting tool for decision-makers that is publicauthorities in choosing the appropriate maintenance mea-sures on time

22 Development of an ANN Prediction Model Today theimplementation of artificial neural networks in order todevelop prediction models becomes a very interestingresearch field which could be used to solve various prob-lems Therefore the idea was to develop such an ANNprediction model which would assist decision-makers indealing with maintenance planning problems of urbanroad infrastructure

In pavement construction according to Croatiannational standards for asphalt pavement design (HRNUC4012) one should take into consideration several layersof materials asphalt materials grain materials stabilized byhydraulic binder unbound grained materials and subgradeMethods for pavement design are divided into two groupsempirical and analytical methods While pavement designempirical methods are used as systematic observations onpavement performance on existing road sections in analyti-cal pavement design mathematical models for calculationare included for determining strain and stresses in pavementlayers that is road deterioration prediction Therefore thecalculated values of stresses and strains are used for estimat-ing pavement life and damage evaluation [22] According toBabić [23] in order to achieve the desired durability of

Envi

ronm

ent

Database Modelbase

Strategicmanagement

level

Tacticalmanagement level

Operationalmanagement level

Dialog

Users

41

2 3

Figure 1 Structure of the decision support concept for urban roadinfrastructure management

3Complexity

pavement construction over time it is necessary to achievethe following

(i) The maximum vertical compressive strain on the topof subgrade does not exceed certain amount

(ii) The horizontal radial stress (strain) at the bottom ofthe cement-bearing layer is less than the allowablestress (strain)

(iii) The horizontal radial stress (strain) at the bottom ofthe asphalt layer is less than the allowable stress(strain)

It is considered that fulfilling the above-stated require-ments protects pavements from premature crack conditionFigure 2 shows the used pavement cross section for themodelling process It is apparent that the observed construc-tion consists of three layers that is asphalt layer unboundgranular material layer and subgrade layer Selected pave-ment structure is under standard load expressed in passagesof ESAL (equivalent single axle load) of 80 kN that is axleloading by 2 wheels on each side with the axle space betweenthem of 35 cm and the axle width of 18m Such road struc-ture is most often used in roads for medium and low trafficloads in the Republic of Croatia

For modelling purposes of development of an ANN pre-diction model only the horizontal radial stress is observedat the bottom of the asphalt layer under the wheel In orderto determine the stress values at the bottom of the asphaltlayer analytically the Circly 60 Pavement Design Software(further Circly 60) is used This software was developed inAustralia several decades ago and since 1987 it has beenan integral part of the Austroads Pavement Design Guidethe standard for road design in Australia and New Zealandas well as a road design worldwide The Circly 60 is a soft-ware package where the rigorous flexible pavement designmethodology concerning both pavement material propertiesand performance models is implemented (httpspavement-sciencecomausoftovercirclycircly6_overview) Materialproperties (Youngrsquos modulus E and Poissonrsquos ratio) loadsand thicknesses of each layer are used as input data whilethe output data is the stress value at the bottom of theasphalt layer

In the second part of the research a diagram (Figure 3) ispresented of the performed tests data collection for themodelling process (1) the division of the total data (2) deter-mination of the ANN model architecture (3) testing of theadopted ANN model (4) analysis of the prediction perfor-mance of an adopted ANN model on independent dataset(5) and application of the adopted ANN model on differenttypes of construction (6)

The ANN model is used for the purpose of achieving asuccessful prediction of horizontal radial stress at the bottomof the asphalt layer The main objective is to produce theANN prediction model based on collected data to test it onan independent dataset subsequently to test the base modelon an extended (independent) dataset and ultimately toanalyze the modelrsquos performance on several pavement struc-tures with variable features

221 Data Collection for the Modeling Process For the needsof the ANNmodel production the used input data are shownin Table 1 In total 560 of the testing combinations areapplied containing variable values of the specific load onthe pavement structure characteristics of the asphalt layer(modulus of elasticity thickness Poissonrsquos ratio and volumebinder content) unbound granular material and subgrade(modulus of elasticity and Poissonrsquos ratio) Circly 60 wasused to determine the stress values (560 testing combina-tions) at the bottom of the asphalt layer (under the wheel)Initial activity is reduced to collecting data from the Circly60 software (560 combinations) where 10 independent vari-ables listed in Table 1 are used as input values The depen-dence of dependent variables (stress) and 10 independentvariables is observed After that the collected data are usedin the process of producing and testing the ANN model

222 Architecture Design of the ANN Model For the pur-poses of this research particularly with the aim of taking intoconsideration the simultaneous impact of multiple variableson the forecasting of asphalt layer stress the feedforwardneural network was used for the ANN prediction modeldevelopment It consists of a minimum of three layers inputhidden and output The number of neurons in the input andoutput layers is defined by a number of selected data whereasthe number of neurons in the hidden layer should be opti-mized to avoid overfitting the model defined as the loss ofpredictive ability [24] Since every layer consists of neuronsthat are connected by the activation function the sigmoidfunction was used The backpropagation algorithm was usedfor the training process The configuration of the appliedneural network is shown in Figure 4

The RapidMiner Studio Version 80 software package isused to develop the ANNmodel In order to design the archi-tecture of the ANN model input data (560) are collectedfrom the Circly 60 Pavement Design Software The total col-lected data are divided into two parts The bigger part (70data in the dataset) is used to design the architecture of the

350 mm

Asphalt

Unbound granularmaterial

Subgrade

Figure 2 Cross section of the observed pavement structure

4 Complexity

ANN model while the remaining (30 of data) is used totest the accepted model Total data are divided into thosewho participate in the process of developing and testingmodels randomly The initial action of the pavement perfor-mance modelling process is to optimize the input parame-ters (momentum learning rate and training cycles) Afterthe optimum (previous) parameters of the methodologyare defined the number of hidden layers and neurons inthe individual layers is determined When the design ofthe ANN model on the training dataset was carried outthe test of the adopted model on an independent dataset(30) is accessed

The optimum combination of the adopted ANN wasthe combination with 1 hidden layer 20 neurons in a singlelayer learning rate 028 and 640 training cycles The adopted

combination allowed the realization of the highest valueof the coefficient of determination (R2= 0992) betweenthe tested and predicted values of tensile stress (for theasphalt layer)

223 Test Cases For the purpose of this research the inputdata were grouped into 3 testing cases (A B and C)

(i) Amdashbase model (determining the ANN model archi-tecture training of the model on a set of 391 inputpatterns and testing of a built-in model on the 167independent dataset)

(ii) Bmdashtesting the A base ANN model on an indepen-dent (extended) dataset (extra 100 test data)

(1) Data collection for themodeling process

(2) e division of the totaldata is on part for making

ANN model (70) and part fortesting of ANN model (30)

(3) Determination of theANN model architecture on

the base dataset (70)

(4) Testing of the adoptedANN model on the

independent dataset (30)

(5) Analysis of the predictionperformance of adopted ANN

model on independent(extended) dataset

(6) Application of theadopted ANN model on

different types ofconstruction

Figure 3 Diagram of the research timeline

Table 1 Input values for the modeling process

Independent variable number Name of input variable Range of used values

1 Specific load on the contact area (05 06 07 and 08MNm2)

2

Asphalt layer

Modulus of elasticity (2000 4000 6000 8000 and 10000MNm2)

3 Thickness (3 6 9 12 15 18 and 21 cm)

4 Poissonrsquos ratio p = 0 355 Volume binder content Bc-v = 13

6

Unbound granular material

Modulus of elasticity Ms = 400MNm2

7 Thickness d = 20ndash80 cm (20 40 60 and 80 cm)

8 Poissonrsquos ratio p = 0 359

SubgradeModulus of elasticity Ms = 60MNm2


Input data x1

Input data xm

Inputlayer

Outputlayer

Hidden layers

Output data

⋮

⋮ ⋮

Figure 4 Configuration of the selected artificial neural network

5Complexity

(iii) Cmdashapplication of the developed ANN model inthe process of forecasting the stresses of asphaltlayers on several different road pavement structures(4 cases)

Following the analysis (Figure 5) an additional test of thebase ANN model (A) on an independent (extended) dataset(B) is performed The extended tested dataset contains anasphalt layer of thickness in the range of 28 to 241 cm theasphalt modulus of elasticity from 1400 to 9700MNm2thickness of unbound granular material from 12 to 96 cmand the specific load on the contact area from 05 to 08MNm2 Part C shows the forecasting success of the adoptedANN model on 4 different pavement structures (indepen-dent data) Consequently the individual relationship ofasphalt layer thickness modulus of elasticity thickness ofunbound granular layer and specific load on the contact areaversus stress is analyzed

3 Results and Discussion

The results of the developed ANN prediction model is shownin Figure 6 in the form of the obtained values of the coeffi-cient of determination (R2) for the observed test cases A Band C It is apparent that a very high coefficient of determina-tion are achieved (from 0965 (B) to 0999 (C4)) The R2

results are consistent with the results of Ghanizadeh andAhadi [25] in their prediction (ANN prediction model) ofthe critical response of flexible pavements

The balance of the paper presents an overview of theobtained results for cases A B and C which also include aview of the testing of the basic ANN model on independentdata Figure 7 shows the linear relationship (y=10219xminus00378) between the real stress values (x) and predicted stressvalues (y) in the case of testing the ANN model on an inde-pendent dataset (30) From the achieved linear relationshipit is apparent that at 6MPa the ANN model will forecast0094MPa higher stress value in comparison to the real stressvalues At 1MPa this difference will be lower by 002MPacompared to the real stress values From the obtained linearrelationship it can be concluded that the adopted ANNmodel achieves a successful forecasting of stress values incomparison to the real results which were not included intothe development process of the ANN model

Figure 8 shows the linear relationship (y=10399xminus00358) between the real stress values (x) and predicted stressvalues (y) for the case of testing the ANN model developedon an independent (extended) dataset From the shown func-tion relationship it is apparent that a lower coefficient ofdetermination of 0965 was achieved with respect to the caseA From the obtained linear relationship it is apparent thatat 4MPa the ANN model will predict 012MPa higher stressvalue in comparison to the real stress values

Table 2 presents the input parameters for 4 differentpavement structures (C1ndash4) where the individual impact ofthe asphalt layer thickness the asphalt elasticity modulusthe thickness of unbound granular material and the specificload on the horizontal radial stress at the bottomof the asphaltlayer (under the wheel) was analyzed For the purposes of the

analysis additional 56 test cases were collected from theCircly 60 software

In combination C1 (Figure 9) a comparison of theasphalt layer thickness and the observed stress for the roadstructure which in its composition has 20 and 90 cm thick-ness of the bearing layer was shown As fixed values specificload on the contact area (07MPa) and modulus of elastici-tymdashasphalt layer (4500MPa) were used Consequently therelationship between the predicted stress and the real values(obtained by using the computer program) is analyzed Theobtained functional relationship clearly shows that theincreasing thickness of the asphalt layer results in anexpected drop in the stress of the asphalt layer This dropin stress is greatest in unbound granular material (20 cm)where it reaches 244MPa between the constructions con-taining 4 cm and 20 cm thick asphalt (stress loss is 09MPain construction with 90 cm thickness of the bearing layer)The largest difference between the predicted (the ANNmodel) and the real stress values in the amount of 025MPa(unbound granular material of 20 cm 4 cm asphalt thickness)was recorded The average coefficient of determination (R2)in combination to C1 amounts to the high 0998

Figure 10 (C2) shows the relationship between the mod-ulus of elasticitymdashasphalt layer and stress in a pavementstructure containing 6 and 18 cm thick asphalt layer In theobserved combination a fixed layer thickness of 50 cm(unbound granular material) and a specific load on the con-tact area of 07MPa are considered As with combination C1the relationship between predicted stress and real values isanalyzed From the obtained functional relationship it isapparent that the growth of the modulus of elasticitymdashas-phalt layer increases its stress as well The increase in stressis greatest in the construction with a thinner asphalt (6 cm)where it amounts to 18MPa between the constructions con-taining the modulus of elasticitymdashasphalt layer of 1400MPaand 9700MPa (stress growth is 05MPa in constructionwith 18 cm thick asphalt layer) The largest differencebetween the predicted (from developed ANN model) andthe real stress values amounts to 0084MPa (modulus ofelasticity 4500MPa 6 cm asphalt thickness) The averagecoefficient of determination (R2) for combinationC2 amountsto high 0996

The following figure (Figure 11) shows the relationshipbetween the thickness of the unbound granular layer andthe asphalt layer stress with variable modulus of elasticity(1400MPa and 9700MPa) As a result a fixed thickness ofasphalt layer (10 cm) and the specific load on the contact areawere applied (05MPa) As an illustration the relationshipbetween the predicted and the real stress of the asphalt isshown From the obtained results it can be seen that withthe increase of the thickness of the unbound granular layerthe stress in asphalt layer is decreased The average coefficientof determination (R2) in combination C3 is high andamounts to 0987 Figure 10 clearly shows a greater differencebetween the real and the predicted stress values on theasphalt layer (1400MPa) The biggest difference in the stressvalues amounts to 018MPa (40 cm thick bearing layer1400MPa modulus of elasticitymdashasphalt layer) Larger devi-ations also lead to a reduction in the individual coefficient of

6 Complexity

determination to the amount of 0958 (for asphalt layer of1400MPa)

The combination C4 (Figure 12) analyzes the effect ofspecific load on the contact area at stress values (the realand predicted values) In the observed combination the valueof the specific load on the contact area ranges from 05 to08MPa It used a 40 cm thick unbound granular layer 4and 16 cm thick asphalt layer and an asphalt layer modulusof elasticity in the amount of 6700MPa The average coeffi-cient of determination (R2) in this combination amounts tohigh 0999 From the obtained results it is apparent that in

the case of 4 cm thick asphalt layer construction there is anincrease in the difference between the real and predictedstress values as the value of the specific load on the contactarea increases As a result this difference is 013MPa at08MPa of the specific load It is also apparent that in the caseof 16 cm thick asphalt construction the ANN predictivemodel does not show any significant change in stress valuesdue to the observed growth in the specific load on the contactarea (this growth at real stress values amounts to a 006MPa)

From the research project carried out it is shown that thepredicted value of stress at the bottom of the asphalt layer is

Asphalt layer thicknessstress

Modulus of elasticitystress

Unbound granular layerthicknessstress

Specific load on the contactareastress

A B C

Figure 5 Test cases

09940992

Coefficient of determination (trainingtesting)

0965 0994

0998

0996

0997

0999

Figure 6 Test results

00

1

2

3

Pred

icte

d str

ess v

alue

(MPa

)

4

5

6

7

1 2 3Real stress value (MPa)

Case A (independent dataset)

R2 = 0993

Case AEquality line

Linear (case A)Linear (equality line)

4 5 6 7

Figure 7 Resultsmdashcase A

7Complexity

successfully achieved by using the developed ANNmodel Aspreviously shown the initial testing process of the deployedANN model was performed on an independent dataset(30 of data in the dataset) which was not used in the train-ing phase of the observed model Having achieved the accept-able result of the forecasting of the dependent variable anindependent (extended) set of testing cases are applied

where the previously deployed ANN model is subsequentlytested Once the trained ANN model is considered as a suc-cessful model the same is applied in the forecasting processon the four different pavement constructions in the furthercourse of the testing process The obtained results confirmthat it is possible to successfully use the developed ANNmodel in the pavement condition forecast methodology

Real stress values (MPa)

Case B (independent dataset)

R2 = 09652

Pred

icte

d str

ess v

alue

s (M

Pa)

Case B

0

1

2

3

4

minus1

Equality line

0 1 2 3 4

Figure 8 Resultsmdashcase B

Table 2 Input valuesmdashcase C

CaseIndependent variable

(x)Dependentvariable (y)

Specific load on thecontact area (MPa)

Unbound granularmaterialmdashthickness (cm)

Asphaltlayermdashthickness

(cm)

Modulus ofelasticitymdashasphalt

(MPa)

C1 Asphalt layer thickness Stress 07 20 and 90 4ndash20 4500

C2Modulus of

elasticitymdashasphaltStress 07 50 6 and 18 1400ndash9700

C3Unbound granularmaterialmdashthickness

Stress 05 20ndash100 10 1400 and 9700

C4Specific load on the

contact areaStress 05ndash08 40 4 and 16 6700

Poissonrsquos ratio p = 0 45 (subgrade) p = 0 35 (unbound granular material) and p = 0 35 (asphalt layer) modulus of elasticitymdashunbound granular material400MPa modulus of elasticitymdashsubgrade 60MPa

2

Unbound granular material (20 cm)Unbound granular material (90 cm)

ANN prediction (20 cm)ANN prediction (90 cm)

0

1

2

Stre

ss (M

Pa)

3

4

4 6 8 10 12Asphalt layer thickness (cm)

14 16 18 20

Figure 9 Resultsmdashcase C1

8 Complexity

After the ANN modelingtesting process it is necessary tocompare the output values obtained with the permissiblevalues In order for the tracked construction to achieve thedesired durability it is also necessary to check that the max-imum vertical compressive strain on the top of subgrade doesnot exceed certain amounts As found by this research thepavement performance forecasting success of the developedANN model also largely depends on the range of input pat-terns used in the modelling process as well as on applicableindependent variables

As such the developed ANN model gives very goodprediction of real stress values at the bottom of the asphaltlayers Compared with analytical results obtained by Circly60 it has very high coefficient of determination for alltested cases which are based on real possibilities of pave-ment structure

As the goal of this research was to design and develop anANN model in order to achieve a successful prediction ofroad deterioration as a tool for maintenance planning activi-ties it can be concluded that such a prediction model basedon ANN is successfully developed Therefore such a modelis part of the proposed decision support concept (DSC) forurban road infrastructure management in its model baseand can be used as a decision support tool in planning activ-ities which occur on the tactical management level

4 Conclusions

The proposed decision support concept and developed ANNmodel show that complex and sensitive decision-makingprocesses such as the ones for urban road infrastructuremaintenance planning can correctly be supported if appro-priate methods and data are properly organized and usedThis paper presents an application of artificial neural net-works in the prediction process of variables concerningurban road maintenance and its implementation in modelbase of decision support concept for urban road infrastruc-ture management The main goal of this research was todesign and develop an ANN model in order to achieve a suc-cessful prediction of road deterioration as a tool for mainte-nance planning activities

Data of the 560 different combinations was obtainedfrom Circly 60 Pavement Design Software and used fortraining testing and validation purposes of the ANN modelThe proposed model shows very good prediction possibilities(lowest R2 = 0987 as highest R2 = 0999) and therefore can beused as a decision support tool in planning maintenanceactivities and be a valuable model in the model base moduleof proposed decision support concept for urban road infra-structure management

Data Availability

The data used to support the findings of this study are avail-able from the corresponding author upon request


The authors declare that there are no conflicts of interestregarding the publication of this paper

References

[1] The World Road Association (PIARC) ldquoThe importance ofroad maintenance The World Road Association (PIARC)rdquo2014 January 2018 httpwwwerfbeimagesImportance_of_road_maintenancepdf

[2] A Deluka-Tibljaš B Karleuša and N Dragičević ldquoReview ofmulticriteria-analysis methods application in decision making

0

1

2

Stre

ss (M

Pa)

1000 2000 3000 50004000 6000 7000 8000 100009000Modulus of elasticitymdashasphalt layer (MPa)

Asphalt layer (6 cm)Asphalt layer (18 cm)



0

1

2

Stre

ss (M

Pa)

3020 5040 8070 9060 100Unbound granular layer thickness (cm)

Asphalt layermdashmodulus of elasticity 9700 MPaAsphalt layermdashmodulus of elasticity 1400 MPaANN (1400 MPa)ANN (9700 MPa)


05

15

25

05 06 07 0804Specific load on the contact area (MPa)


ANN (4 cm)ANN (16 cm)

Stre

ss (M

Pa)


9Complexity

about transport infrastructurerdquo Građevinar vol 65 no 7pp 619ndash631 2013

[3] T Hanak I Marović and S Pavlović ldquoPreliminary identifica-tion of residential environment assessment indicators for sus-tainable modelling of urban areasrdquo International Journal forEngineering Modelling vol 27 no 1-2 pp 61ndash68 2014

[4] I Marović Decision Support System in Real Estate Value Man-agement [PhD Thesis] University of Zagreb Faculty of CivilEngineering Zagreb Croatia 2013

[5] N Jajac I Marović and T Hanak ldquoDecision support for man-agement of urban transport projectsrdquo Građevinar vol 67no 2 pp 131ndash141 2015

[6] N Jajac S Knezić and I Marović ldquoDecision support systemto urban infrastructure maintenance managementrdquo Organiza-tion Technology amp Managament in Construction an Interna-tional Journal vol 1 no 2 pp 72ndash79 2009

[7] B Karleuša A Deluka-Tibljaš and M Benigar ldquoPossibility forimplementation of multicriteria optimalisation in the trafficplanning and designrdquo Suvremeni promet vol 23 no 1-2pp 104ndash107 2003

[8] D Moazami H Behbahani and R Muniandy ldquoPavementrehabilitation and maintenance prioritization of urban roadsusing fuzzy logicrdquo Expert Systems with Applications vol 38no 10 pp 12869ndash12879 2011

[9] F Wang Z Zhang and R Machemehl ldquoDecision-makingproblem for managing pavement maintenance and rehabilita-tion projectsrdquo Transportation Research Record Journal of theTransportation Research Board vol 1853 pp 21ndash28 2003

[10] K Zimmerman and D Peshkin ldquoIssues in integrating pave-ment management and preventive maintenancerdquo Transporta-tion Research Record Journal of the Transportation ResearchBoard vol 1889 pp 13ndash20 2004

[11] H Zhang G A Keoleian and M D Lepech ldquoNetwork-levelpavement asset management system integrated with life-cycleanalysis and life-cycle optimizationrdquo Journal of InfrastructureSystems vol 19 no 1 pp 99ndash107 2013

[12] R Denysiuk A V Moreira J C Matos J R M Oliveira andA Santos ldquoTwo-stage multiobjective optimization of mainte-nance scheduling for pavementsrdquo Journal of InfrastructureSystems vol 23 no 3 article 04017001 2017

[13] M A Abo-Hashema and E A Sharaf ldquoDevelopment of main-tenance decision model for flexible pavementsrdquo InternationalJournal of Pavement Engineering vol 10 no 3 pp 173ndash1872009

[14] N Zavrtanik J Prosen M Tušar and G Turk ldquoThe use ofartificial neural networks for modeling air void content inaggregate mixturerdquo Automation in Construction vol 63pp 155ndash161 2016

[15] D Singh M Zaman and S Commuri ldquoArtificial neural net-work modeling for dynamic modulus of hot mix asphalt usingaggregate shape propertiesrdquo Journal of Materials in Civil Engi-neering vol 25 no 1 pp 54ndash62 2013

[16] I Androjić and I Marović ldquoDevelopment of artificial neuralnetwork and multiple linear regression models in the predic-tion process of the hot mix asphalt propertiesrdquo Canadian Jour-nal of Civil Engineering vol 44 no 12 pp 994ndash1004 2017

[17] M Bielli ldquoA DSS approach to urban traffic managementrdquoEuropean Journal of Operational Research vol 61 no 1-2pp 106ndash113 1992

[18] A Quintero D Konare and S Pierre ldquoPrototyping anintelligent decision support system for improving urban

infrastructures managementrdquo European Journal of Opera-tional Research vol 162 no 3 pp 654ndash672 2005

[19] N Jajac Design of Decision Support Systems in the Manage-ment of Infrastructure Systems of the Urban Environment[MS thesis] University of Split Faculty of Economics SplitCroatia 2007

[20] I Marović I Završki and N Jajac ldquoRanking zones model ndash amulticriterial approach to the spatial management of urbanareasrdquo Croatian Operational Research Review vol 6 no 1pp 91ndash103 2015

[21] E Turban Decision Support and Expert Systems ManagementSupport Systems Macmillan Publishing Company New YorkNY USA 1993

[22] N Vukobratović I Barišić and S Dimter ldquoInfluence ofmaterial characteristics on pavement design by analytical andempirical methodsrdquo e-GFOS vol 14 pp 8ndash19 2017

[23] B Babić Pavement Design Croatian Association of CivilEngineers Zagreb 1997

[24] S O Haykin Neural Networks and Learning MachinesPearson Education Upper Saddle River NJ USA 2009

[25] A R Ghanizadeh andM Reza Ahadi ldquoApplication of artificialneural networks for analysis of flexible pavements under staticloading of standard axlerdquo International Journal of Transporta-tion Engineering vol 3 no 1 pp 31ndash42 2016

10 Complexity

Research ArticleANN Based Approach for Estimation of Construction Costs ofSports Fields

MichaB Juszczyk Agnieszka LeVniak and Krzysztof Zima

Faculty of Civil Engineering Cracow University of Technology 24 Warszawska St 31-155 Cracow Poland

Correspondence should be addressed to Michał Juszczyk mjuszczykizwbitpkedupl

Received 29 September 2017 Accepted 13 February 2018 Published 14 March 2018

Academic Editor Andrej Soltesz

Copyright copy 2018 Michał Juszczyk et alThis is an open access article distributed under the Creative CommonsAttribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Cost estimates are essential for the success of construction projects Neural networks as the tools of artificial intelligence offer asignificant potential in this field Applying neural networks however requires respective studies due to the specifics of differentkinds of facilities This paper presents the proposal of an approach to the estimation of construction costs of sports fields which isbased on neural networks The general applicability of artificial neural networks in the formulated problem with cost estimation isinvestigated An applicability of multilayer perceptron networks is confirmed by the results of the initial training of a set of variousartificial neural networks Moreover one network was tailored for mapping a relationship between the total cost of constructionworks and the selected cost predictors which are characteristic of sports fields Its prediction quality and accuracy were assessedpositively The research results legitimatize the proposed approach

1 Introduction

The results presented in this paper are part of a broadresearch in which the authors participate aiming to developtools for fast cost estimates dedicated to the constructionindustry The main aim of this paper is to present the resultsof the investigations on the applicability of artificial neuralnetworks (ANNs) in the problem of estimating the total costof construction works in the case of sports fields as specificfacilities The authors propose herein a new approach basedon ANNs for estimating construction costs of sports fields

11 Cost Estimation in Construction Projects Cost estimationis a key issue in construction projects Both underestimationand overestimation of costs may lead to a failure of aconstruction projectTheuse of different tools and techniquesin the whole project life cycle should provide informationabout costs to the participants of the project and support acomplex decision-making process In general cost estimatingmethods can be classified as follows [1 2]

(i) Qualitative cost estimating

(1) Cost estimating based on heuristic methods(2) Cost estimating based on expert judgments

(ii) Quantitative cost estimating

(1) Cost estimating based on statistical methods(2) Cost estimating based on parametric methods(3) Cost estimating based on nonparametric meth-

ods(4) Cost estimating based on analogouscompara-

tive methods(5) Cost estimating based on analytical methods

The expectations of the construction industry are toshorten the time necessary to predict costs whilst onthe other hand the estimates must be reliable and accu-rate enough There are worldwide publications in whichthe authors report the research results which respond tothese expectations The examples of the use of a regres-sion analysis (based on both parametric and nonpara-metric methods) are as follows application of multivari-ate regression to predict accuracy of cost estimates onthe early stage of construction projects [3] implementa-tion of linear regression analysis methods to predict thecost of raising buildings in the UK [4] proposal anddiscussion of the construction cost estimation methodwhich combines bootstrap and regression techniques [5]


2 Complexity

and application of boosting regression trees in preliminarycost estimates for school building projects in Korea [6]Another mathematical tool for which some examples can begiven is fuzzy logic for example implementation of fuzzylogic for parametric cost estimation in construction buildingprojects in Gaza Strip [7] or proposal and presentation of afuzzy risk assessment model for estimating a cost overrunrisk rating [8] Case based reasoning (CBR) is also anapproachwhich can be found in the publications dealingwiththe construction cost issue for example implementation ofthe CBR method improved by analytical hierarchy process(AHP) for the purposes of cost estimation of residentialbuildings in Korea [9] or the use of the case based reasoningin cost estimation of adapting military barracks also in Korea[10] The examples of the publications which report anddiscuss the applications of artificial neural networks in thefield of cost estimation and cost analyses in the constructionprocess are presented in the next subsection

12 Artificial Neural Networks Cost Estimation inConstructionProjects Artificial neural networks (ANNs) can be definedas mathematical structures and their implementations (bothhardware and software) whose mode of action is based onand inspired by nervous systems observed in nature In otherwords ANNs are tools of artificial intelligence which have theability to model data relationships with no need to assume apriori the equations or formulaswhich bind the variablesThenetworks come inwide variety depending on their structuresway of processing signals and applications The theory inthis subject is widely presented in the literature (eg [11ndash15]) Main applications of ANN can be mentioned as follows(cf eg [11 12 15]) prediction approximation controlassociation classification and pattern recognition associat-ing data data analysis signal filtering and optimizationANNs features whichmake thembeneficial in cost estimatingproblems (in particular for cost estimating in construction)are as follows

(i) Applicability in regression problems where the rela-tionships between the dependent and many indepen-dent variables are difficult to investigate

(ii) Ability to gain knowledge in the automated trainingprocess

(iii) Ability to build and store the knowledge on the basisof the collected training patterns (real-life examples)

(iv) Ability of knowledge generalization predictions canbe made for the data which have not been presentedto the ANNs during a training process

Some examples of ANN applications reported for a rangeof cost estimating and cost analyses in construction arereplication of past cost trends in highway construction andestimation of future costs trends in this field in the state ofLouisiana USA [16] computation of the whole life cost ofconstruction with the use of the concept of cost significantitems in Australia [17] prediction of the total structural costof construction projects in the Philippines [18] estimationof site overhead costs in the dam project in Egypt [19]

prediction of the cost of a road project completion on thebasis of bidding data in New Jersey USA [20] and costestimation of building structural systems in Turkey [21] Theauthors of this paper also have their contribution in studies onthe use of ANN in cost estimation problems in constructionIn some previous works the authors presented the ANNapplications for conceptual cost estimation of residentialbuildings in Poland [22ndash24] and estimation of overhead costin construction projects in Poland [25 26]

13 Justification for Research It needs to be emphasizedthat despite a number of publications reporting researchprojects on the use of artificial neural networks in costanalyses and cost estimation in construction each of theproblems is specific and unique Each of such problemsrequires an individual approach and investigation due todistinct conditions determinants and factors that influencethe costs of construction projects An individual approach tocost estimation in construction is primarily due to specificityof the facilities including sports fields The costs of a sportfield are significant not only for the construction stage butalso later in terms of its maintenance The decisions madeabout the size functionality and quality are crucial for thefuture use and operational management of sport fields Thesuccess in investigation of ANNs applicability in the problemwill allow proposing a new approach for estimation of theconstruction cost of sport fields The new approach basedon the advantages offered by neural networks will allowpredicting the total construction cost of sport fields muchfaster than with traditional methods moreover it will givethe possibility of checking many variants and their influenceon the cost in a very short time

2 Formulation of the Problem andResearch Framework

21 General Assumptions Thegeneral aimof the researchwasto develop a model that supports the process of estimatingconstruction costs of sports fields The authors decided toinvestigate implementation of ANNs for the purpose ofmapping multidimensional space of cost predictors into aone-dimensional space of construction costs In a formalnotation the problem can be defined generally as follows

119891 119883 997888rarr 119884 (1)

where 119891 is sought-for function of several variables 119883 isinput of the function 119891 which consists of vectors 119909 =[1199091 1199092 119909푛] where variables 1199091 1199092 119909푛 represent costpredictors characteristic of sports fields as constructionobjects and 119884 is a set of values which represent constructioncosts of sports fields

In the statistical sense the problem comes down tosolving a regression problem and estimating of a relationship119891 between the cost predictors being independent variablesbelonging to the set 119883 and constructions cost of a sportsfield being dependent variable belonging to the set 119884According to the methodology in cost estimating basedon statistical methods one can distinguish between two

Complexity 3

Analysis of the problem(formulation of the problem studying available information about completed construction projects on the sports fields establishing general assumptions for the problem and preselection of the cost predictors)

Collection of the data relevant for the cost estimation of sports fields

(collecting and ordering the information on the basis of analysis of completed construction projects on sports fields)

Analysis of the collected data(elimination of the outliers analysis of the correlation significance between the cost and preselected cost predictors)

Are the initial results satisfying Continue research

Yes

No

Further training of selected neural networks analysis of the training results assessment of neural networks performance and applicability for the problem

Selection of the final set ofcost predictors

Initial training of the neural networks

(preliminary assessment of the applicability of the neural networks)

Figure 1 Scheme of the research framework (source own study)

main approaches estimating based on parametric methodsand estimating based on nonparametric methods (cf [12 25]) Both methods rely on the real-life data that isrepresentative samples of cost predictors values and relatedconstruction costs values In the case of the use of parametricmethods function 119891 is assumed a priori and the structuralparameters of the model are estimated On the other handnonparametricmethods are based on fitting the function119891 tothe data According to the assumptions made for the researchpresented in this paper the sought-for functionwas supposedto be implemented implicitly by ANN

A general framework of the adopted research strategy isdepicted in Figure 1

22 Characteristics of Sports Fields Covered by the ResearchSports fields are facilities for which some types of works areusually repeated during the construction stage The maintypes of works that can be listed are

(i) geodetic surveying(ii) earthworks (topsoil stripping trenching compacting

of the natural subgrade etc)(iii) works on subgrade preparation for the sports field

surface

(iv) works on sports fields surface (usually surfaces areeither natural or synthetic grass)

(v) assembly of fixtures and in-ground furnishings (egfootballhandball gates basketball goal systems vol-ley ball or tennis poles and nets)

(vi) works on fencing and ball-nets installation

(vii) minor road works and works on sidewalks

(viii) landscape works and arranging green areas aroundthe sports fields

In the course of the research a number of completedprojects on sports fields in Poland were investigated Boththe fields dedicated to one discipline and multifunctionalfields were taken into account The facilities subject to theanalysis differed in size of the playing area arranged areafor communication arranged green area and fencing Thesurfaces were of two types either natural or synthetic grassIt must be stressed here that the quality expectations forsurfaces varied significantly and played an important role inthe construction costs The completed facilities are locatedall over Poland both in the urban areas (in cities of differentsizes) and outside the urban areas (in the villages)

4 Complexity

Table 1 Characteristics of dependent variables and independent variables considered initially to be used in the course of a regression analysis(source own study)

Description of the variable Variable type ValuesTotal cost of construction works Quantitative Cost given in thousands of PLNPlaying area of the sports field Quantitative Surface area measured in m2

Location of the facility QuantitativeUrban area (big cities medium citiessmall cities) or outside the urban area

(villages)

Number of sport functions Quantitative Number of sports that can be played inthe field

Type of the playing field surface Categorical Natural or artificial grass

Quality standard of the playing fieldrsquossurface Categorical

Quality standard assessed according tothe available information in the tender

documentationBall stop netrsquos surface Quantitative Surface area measured in m2

Arranged area for communication Quantitative Surface area measured in m2

Fencing length Quantitative Length measured in mArranged green area Quantitative Surface area measured in m2

3 Variables Analysis

31 Preselection of Variables As the problem was formallyexpressed and the assumption about the use of ANNs wasmade the authors focused on the analyses which allowedthem to preselect cost predictors The preselection waspreceded by studying both technical and cost aspects ofthe construction projects on sports fields This stage of theresearch allowed collecting the necessary background knowl-edge about the nature of sports fields as specific constructionobjects with their characteristic elements range and sequenceof construction works which must be completed and theclientsrsquo quality expectations

In the next step 129 construction projects on sportsfields that were completed in Poland in recent years wereinvestigated For the purposes of fast cost analysis the authorspreselected the following data to be the variables of thesought-for relationship

(i) Total cost of construction works as a dependentvariable

(ii) Playing area of a sports field location of a facilitynumber of sport functions the type of the playingfieldrsquos surface (natural or artificial) quality standardof the playing fieldrsquos surface ball stop netrsquos surfacearranged area for communication fencingrsquos lengthand arranged greenery area as independent variables

The criteria for such preselection were the availability ofthe data in the investigated tender documents and ensuringenough simplicity of the developed model due to which thepotential client would be able to formulate the expectationsabout the sport field to be ordered by specifying values forpotential cost predictors in the early stage of the project

Most of the mentioned variables were of a quantitativetype In the case of the location of the facility type of theplaying fieldrsquos surface and quality standard of the playingfieldrsquos surface only descriptive information was available the

three variables were of the categorical type Table 1 presentssynthetically the characteristics of all of the variables thatwere preselected in the course of the analysis of the problem

The next step included data collection and scaling cate-gorical variables In the case of three of the variables (namelylocation of the facility type of the playing field surface andquality standard of the playing fieldrsquos surface) categoricalvalues were replaced by numerical values The studies of theproblem analyses of the number of completed constructionprojects on sports fields and especially the analyses ofconstruction works costs brought the conclusions of how thecategorical values of the three variables are associatedwith thecosts of construction works Table 2 explains how the changeof the categorical values stimulates the costs of constructionworks in general The observation made it possible to orderthe values and transform them into numbers in the rangefrom 01 to 09

Categorical values for location of the facility have takennumerical values as follows urban area 09 for big cities(population over 100000) 066 for medium cities (pop-ulation between 20000 and 100000) and 033 for smallcities (population below 20000) outside the urban area01 for villages In the case of the type of the playing fieldsurface artificial grass has taken the value of 09 and naturalgrass has taken the value of 01 Finally depending on theclientrsquos expectations and specifications available in the tenderdocuments the descriptions of the demands for qualitystandard of the playing fieldrsquos surface took values from therange between 01 and 09

The studies of tender documents for public constructionprojects where completion of sports fields was the subjectmatter of the contract allowed for collecting the data for 129projects The data were collected for projects completed inthe last four years all over Poland The collected informationwas ordered in the database After the analysis of outliersthe authors decided to reject some extreme cases for whichthe total construction cost was unusually high or unusually

Complexity 5

Table 2 General relationship between the three categorical variables and cost (source own study)

Location of a facility Type of the playing field surface Quality standard of the playing fieldrsquos surface CostUrban area big cities Artificial grass High demands HigherUrban area medium cities Moderate demandsUrban area small cities Natural grass LowerOutside the urban area villages Low demands

Table 3 Significance of correlations between the variables (source own study)

Preselected cost predictors

Is the correlation between the dependentvariable total cost of construction worksand independent variable significant for119901value lt 005

Variablersquos symbol (for accepted variablesonly)

Playing area of the sports field Yes 1199091Location of the sports ground No -Number of sport functions No -Type of the playing field surface Yes 1199092Quality standard of the playing fieldrsquos surface Yes 1199093Ball stop netrsquos surface Yes 1199094Arranged area for communication Yes 1199095Fencing length Yes 1199096Arranged greenery area Yes 1199097low After the elimination of outliers the data for 115 projectsremained

32 Selection of the Final Set of Variables Further analysisincluded the investigation of the significance of correlationsbetween the dependent variable and all of the initially consid-ered independent variables preselected cost predictors Thesignificance of correlations for 119901-value lt 005 was assessedThe results of this step are synthetically presented in Table 3

As the correlations for the two of preselected costpredictors (namely location of the sports ground and thenumber of sport functions) appeared to be insignificant theywere rejected and no longer taken into account as the costpredictors

Table 4 presents ten exemplary records with the specificnumerical values of the dependent variable 119910 (total costof construction works) and independent variables 119909푗 (costpredictors) accepted for the model

Table 5 presents descriptive statistics for the variablesaccepted for the model Average minimum and maximumvalues are presented for each of the variables as well as thestandard deviation

It is noteworthy that minimum value namely 000 forthe variables 1199094 1199095 1199096 and 1199097 corresponds with the factthat in case of certain sports fields elements such as ballstop nets arranged area for communication fencing andarranged greenery have not been included in a projectrsquos scope(cf Table 4) Moreover there is some regularity in Table 4whichmanifests in the distribution of average values closer tominimum values than to maximum values This is due to thefact that the number of small-sized and medium-sized sportsfields in the database was relatively greater than the numberof large-sized ones This can be explained by a general rule

valid for all kinds of constructionThe number of small-sizedand medium-sized facilities of all types either newly built orexisting is always greater than those large-sized

The database records (whose number equalled 115) wereused as training patterns 119901 for ANNs in the course of theresearch

The values of the variables were scaled automaticallybefore and after each of the ANNrsquos training This was donedue to the functionalities of the ANNrsquos software simulatorused in the course of the research The variables were scaledlinearly to the range of values appropriate for activation func-tions employed for certain investigated ANN The resultsespecially ANNsrsquo training errors presented further in thepaper are given as original not scaled values

4 Initial Training of Neural Networks

After the selection of independent variables a formal nota-tion of the relationship in the statistical sense can be given asfollows

119910 = 119891 (1199091 1199092 1199093 1199094 1199095 1199096 1199097) + 120576 (2)

Consequently a prediction can be formally made

119910 = 119891 (1199091 1199092 1199093 1199094 1199095 1199096 1199097) (3)

where 119910 is dependent variable total cost of constructionworks as observed in real life 119910 is predicted total cost ofconstruction works 119891 is sought-for function implicit rela-tionship implemented by ANN 1199091 1199092 1199093 1199094 1199095 1199096 1199097 areindependent variables selected cost predictors as presentedin Tables 3 and 4 and 120576 are random deviations (errors) forwhich 119864(120576) = 0

6 Complexity

Table 4 Exemplary records of the database including training patterns (source own study)

119901 119910 1199091 1199092 1199093 1199094 1199095 1199096 11990975 5658 968 09 06 00 1968 6025 0013 1359 3292 09 05 6000 21420 00 0023 4276 1860 09 03 11160 780 375 1000037 4895 1860 01 03 2400 1000 00 0046 3230 800 09 05 1920 1818 1397 0059 1813 800 01 03 720 00 00 400067 19723 4131 09 05 3960 10960 2071 2586482 1611 1650 01 01 3000 00 00 0094 2500 1470 01 02 3445 939 385 00101 8005 1104 09 08 2959 14691 933 3746

Table 5 Descriptive statistics for the modelsrsquo variables (source own study)

Variablersquos symbol Average value Minimum value Maximum value Standard deviationDependent variable 119910 45756 3330 259250 37314

Independent variables

1199091 133379 27500 560000 788491199092 059 010 090 0271199093 049 010 090 0161199094 30026 000 221200 345271199095 19316 000 214200 325841199096 10524 000 60250 121131199097 32429 000 300000 60326

The aim of this stage of the research namely the initialtraining of ANNs was to assess their applicability to theproblem in general and to take a decisionwhether to continuethe research or not A variety of feed forward ANNs weretrained in the automatic mode The overall number ofnetworks equalled 200 the authors took into account 100multilayer perceptron (MLP) networks and 100 radial basisfunction (RBF) networks as the types appropriate for theregression analysis and suitable for the formulated problem

The main criteria to assess the applicability of the ANNswere the quality of predictions made by trained networksand the errors The measure for quality of predictions wasPearsonrsquos correlation coefficient 119877(119910 119910) between that in reallife and that predicted by networks values of the independentvariable the total construction cost whereas the measures oferror was the root mean squared error (RMSE)

119877 (119910 119910) = cov (119910 119910)120590푦120590푦

119877119872119878119864 = radic 1119899sum푝 (119910푝 minus 119910푝)2

(4)

where cov(119910 119910) is covariance between 119910 and 119910 120590푦 is standarddeviation for 119910 and 120590푦 is standard deviation for 119910

In the case of RBF networks both the quality of predic-tions and errors were so dissatisfying that the authors decidedto focus on the MLP networks only The results for the MLPnetworks were satisfying they are presented syntheticallybelow in Figure 2 and in Table 5 and discussed briefly Themain assumptions for this stage were as follows

(i) From the database of training patterns learning (119871)validating (119881) and testing (119879) subsets were randomlydrawn 10 times

(ii) The overall number of available training patterns wasdivided into the three subsets in relation 119871119881119879 =602020

(iii) For each drawing 10 different networks were trained(iv) The networks varied in the number of neurons in the

hidden layer119867(v) Distinct activation functions such as linear sigmoid

hyperbolic tangent and exponential were applied inthe neurons of a hidden and output layer

Learning and validating that is 119871 and119881 subsets were used inthe course of training process The third subset 119879 was usedfor testing purposes after completing the training process asan additional check of the generalization capabilities (cf eg[11]) Number of neurons in the hidden layer119867 was assessedaccording to the following equation [27] and inequality [28]

119867 asymp radic119873119872 (5)

where 119873 is number of neurons in the input layer and 119872 isnumber of neurons in the output layer

NNP = NNW + NNB lt 119871 (6)

whereNNP is number ofANNrsquos parametersNNWis numberof ANNrsquos weights NNB is number of ANNrsquos biases and 119871 iscardinality of a learning subset

Complexity 7

Table 6 Summary of the initial training of ANNs for MLP networks (source own study)

Subset 119877 (119910 119910) RMSEAverage Standard deviation 119877 gt 09 Average Standard deviation

119871 0901 0240 821 1455 1117119881 0879 0228 810 1088 578119879 0881 0233 805 1269 833

minus100

minus080

minus060

minus040

minus020

000

020

040

060

080

100

minus100minus080minus060minus040minus020 000 020 040 060 080 100

RT(y y)

RL(y y)

(a)

0

100

200

300

400

500

0 100 200 300 400 500

２－

３T

２－３L

(b)

Figure 2 Quality end errors of ANNs after the initial training phase (a) scatter diagram of Pearsonrsquos correlation coefficients and (b) scatterdiagram of errors (RMSE) 119871 and 119879 stand for learning and testing subsets accordingly (source own study)

From (5)119867 asymp 2646 According to the assumptions aboutthe cardinality of119871 subset and from inequality (6) NNP lt 69Compromising these two conditions the number of neuronsin the hidden layer119867 varied between 2 and 6

The overall results in terms of the quality of predictionsand errors are presented in Figures 2 and 3

Part (a) of Figures 2 and 3 depicts scatter plots of Pearsonrsquoscorrelation coefficients calculated for each network after thetraining process Figure 2 shows coefficients for learning andtesting subsets whereas Figure 3 shows the coefficients forlearning and validating subsets In part (b) of both figuresone can see scatter plots of errors (namely RMSE) Figure 2presents the scatter plot for learning and testing subsets andFigure 3 presents the scatter plot for learning and validatingsubsets

Table 6 presents the summary of the initial training of 100MLPnetworksThe average and standard deviation of119877(119910 119910)as well as percentage of the cases for which 119877(119910 119910) is greaterthan 09 are presented in the table Additionally the averageand standard deviation for RMSE errors are given All valueswere calculated for learning validating and testing subsetsseparately

As can be seen both in Figures 2 and 3 and in Table 6the correlations for most of the cases are very high For morethan 80 of networks 119877(119910 119910) was greater than 09 in case oflearning validating and testingThere are evident clusters ofpoints in Figures 2(a) and 3(a) which represent networks withthe potential of the acceptable or good quality of prediction

There are only few cases outside the clusters for which thecorrelations coefficients are very low due to the failure of thetraining process RMSE errors are in the acceptable range atthis stage

This stage of the research confirmed the general appli-cability of ANNs to the investigated problem The decisionwas to continue the research Moreover it allowed choosinga group of MLP networks to be trained in the next stage

5 Results of Neural Networks Training in theClosing Phase of the Research

With respect to the initial training results a group of 5networks was chosen for the closing phase of the researchThe details including networksrsquo structures and activationfunctions are given in Table 7 (All of the networks consistedof 7 neurons in the input the number of neurons rangingfrom 2 to 5 in one hidden layer and one neuron in theoutput layer All of the networks were trained with the useof BroydenndashFletcherndashGoldfarbndashShanno (BFGS) algorithm)

Assumptions for the networks training and testing weredifferent than in the initial stage From the set of 115 trainingpatterns the testing subset 119879 was chosen randomly inthe beginning This subset remained unchanged for all ofthe networks and it was used to assess the generalizationcapabilities after all of the training processes Cardinality of119879subset equalled 10 of the total number of training patterns

8 Complexity

000

020

040

060

080

100

000 020 040 060 080 100

minus100

minus080

minus060

minus040

minus020minus100minus080minus060minus040minus020

RV(y y)

RL(y y)

(a)

0

100

200

300

400

500

0 100 200 300 400 500

２－

３V

２－３L

(b)

Figure 3 Quality end errors of ANNs after the initial training phase (a) scatter diagram of Pearsonrsquos correlation coefficients and (b) scatterdiagram of errors (RMSE) 119871 and 119881 stand for learning and validating subsets accordingly (source own study)

Table 7 Details of the selected ANNs for further training (source own study)

ANN Number of neurons inthe hidden layer

Activation functionhidden layer

Activation functionoutput layer Training algorithm

MLPe-l 7-2-1 2 Exponential Linear BFGSMLPe-ht 7-2-1 2 Exponential Hyperbolic tangent BFGSMLPe-l 7-3-1 3 Exponential Linear BFGSMLPs-l 7-5-1 5 Sigmoid Linear BFGSMLPht-e 7-5-1 5 Hyperbolic tangent Exponential BFGS

The remaining patterns have been involved in the 10-fold cross-validation of the networks (cf [11]) The patternsavailable for the training process after the sampling of the119879 subset have been divided into the 10-fold complementarylearning and validating subsets The relation of the subsetswas 119871119881 = 9010 accordingly The sum of squared errorsof prediction (SSE) was used as the error function in thecourse of training

SSE = sum푝isin퐿

(119910푝 minus 119910푝)2 (7)

The performance of ANNs was assessed in general in thelight of correlation between real-life and predicted valuesRMSE errors of the prediction (as in the stage of initialtraining) Table 8 presents a summary of the training resultsfor the five chosen networks after the training process basedon 10-fold cross-validation approachThe results are given interms of the networksrsquo performance and errors To synthesizeand assess the performance and stability of training of eachnetwork maximum average and minimum as well as thedispersion of the correlation coefficients 119877(119910 119910) betweenreal-life and predicted values were calculated Errors namelyRMSE values are presented in the same manner Both

correlations and errors are given separately for 119871 119881 and 119879subsets The analysis of the results made it possible to choosefinally one of the five networksThe most stable performancewas observed for the MLPs-l 7-5-1

Figure 4 depicts a scatter plot of the points that representpairs (119910푝 119910푝) for the finally chosen network MLPs-l 7-5-1Real-life values119910 are set together with predicted values119910Thepoints represent training results for learning and validatingsubsets as well as testing results for testing subset In thelegend of the chart apart from the letters 119871 119881 and 119879 thatexplain membership of the certain pattern to the learningvalidating or testing subset accordingly the numbers from1 to 10 are given next to each letter These numbers revealthe 119896th fold of the cross-validation process In Figure 4 onecan see that the points in the scatter plot are decomposedalong a perfect fit line The deviations are in the acceptablerange In respect of the analysis of 119877(119910 119910) RMSE errorsand distribution of points in the scatter plot the results aresatisfactory

Two more criteria relating to the accuracy of costestimation were also specified for assessment of the selectednetworkMLPs-l 7-5-1The accuracy of estimates was assessedwith the use of three error measures mean absolute percent-age error (MAPE) absolute percentage error calculated for

Complexity 9

Table 8 Results of the selected ANNs training (source own study)

ANN 119877 (119910 119910) RMSE119871 119881 119879 119871 119881 119879MLPe-l 7-2-1

Max 0992 0997 0997 92043 111521 68006Average 0985 0982 0993 66416 63681 41286Min 0973 0948 0985 49572 20138 26656Standard deviation 0005 0015 0004 11620 24243 12338

MLPe-ht 7-2-1Max 0994 0994 0991 150544 376102 79538Average 0978 0979 0983 73351 90999 57609Min 0933 0933 0979 43473 40899 35816Standard deviation 0022 0018 0003 31521 95430 13082

MLPe-l 7-3-1Max 0994 0994 0991 150544 376102 79538Average 0978 0979 0983 73351 90999 57609Min 0933 0933 0979 43473 40899 35816Standard deviation 0022 0018 0003 31521 95430 13082119872119871119875s-l 7-5-1Max 0997 0994 0996 65226 96700 71010Average 0992 0983 0991 46222 59770 39252Min 0986 0949 0980 30333 29581 19582Standard deviation 0004 0015 0008 12048 17682 18116

MLPht-e 7-5-1Max 0996 0995 0996 129738 211452 83939Average 0979 0980 0980 77525 72796 50383Min 0946 0957 0952 32191 41891 27691Standard deviation 0013 0013 0014 27375 49569 18072

Table 9 MAPE and PEmax errors calculated for the MLPs-l 7-5-1 (source own study)

MLPs-l 7-5-1 119871 119881 119879MAPE

Max 1876 2513 2120Average 1268 1443 997Min 749 554 560Standard deviation 349 566 444

PEmax


each pattern (PE푝) and maximum absolute percentage error(PEmax)

MAPE = 100119899 sum푝100381610038161003816100381610038161003816100381610038161003816119910푝 minus 119910푝119910푝

100381610038161003816100381610038161003816100381610038161003816

PE푝 = 100381610038161003816100381610038161003816100381610038161003816119910푝 minus 119910푝119910푝

100381610038161003816100381610038161003816100381610038161003816 100PEmax = max (PE푝)

(8)

Table 9 shows the maximum average and minimum MAPEand PEmax errors as well as the standard deviations ofthese errors after the 10-fold cross-validation for the selectednetwork MLPs-l 7-5-1 The errors are given for the learningvalidating and testing that is 119871 119881 and 119879 subsets respec-tively

MAPE and PEmax errors have been carefully investigatedfor the selected network in all of 10 cases of cross-validationtraining and testing In respect of average MAPE errors theresults were satisfying Average MAPE errors were expectedto be smaller than 15 for learning validating and testing

10 Complexity

0

500

1000

1500

2000

2500

3000

0 500 1000 1500 2000 2500 3000y

L1V1T1L2V2T2L3V3T3L4

V4T4L5V5T5L6V6T6L7V7

T7L8V8T8L9V9T9L10V10T10

y

Figure 4 Scatter plot of 119910 and 119910 values for the selected MLPs-l 7-5-1network (source own study)

subsets This objective was achieved In case of PEmax errorsthe authors expected the values lower than 30 Table 9reveals that the PEmax errors are greater than 30 This facthas prompted the authors to examine PE푝 errors Analysisof the distribution of PE푝 errors revealed that most of theseerrors were smaller than 30 and only few values in each ofthe cross-validation folds values exceeded 30

A thorough analysis of the selected network namelyMLPs-l 7-5-1 in terms of the training results performanceand errors allowed selecting this network to implementimplicitly the sought-for relationship 119891 In conclusion theselected MLPs-l 7-5-1 may be proposed as the tool whichsupports cost estimation of constructionworks in the projectsrelated to sports fields

6 Summary and Conclusion

In this paper the authors presented their investigations onapplicability of ANNs in the problem of estimating the totalcost of construction works for sports fields The researchallowed the authors to confirm assumptions about the generalapplicability of the MLP type networks as the tool which hasthe potential of mapping the relationship between the totalcost of construction works and selected cost predictors whichare characteristic of sports fields On the other hand the RBFtype networks appeared to not be suitable for this particularcost estimation problem

Apart from the general conclusions about the applicabil-ity of ANNs one type of network tailored for this problemwas selected from a broad set of various MLP networks The

analysis of the results indicates a satisfactory performance ofthe selected network in terms of correlations between the realcost and cost predictions The level of the MAPE and RMSEerrors is acceptable For now the proposed approach allowsthe following

(i) Estimating cost of construction works for a couple ofvariants of sports fields in a very short time

(ii) Supporting the decisions made by the client beingaware of the range of the cost estimation accuracy

The obtained results encourage continuation of the inves-tigations which will aim to improve the model especially tolower the PEmax errors The authors intend to collect addi-tional data which will enable them to exceed the number oftraining patterns Moreover the authors intend to investigatein the near future more complex tools based on ANNs suchas committees of neural networks



Acknowledgments

The authors use the STATISTICA software in their researchtherefore some of the presented assumptions for neuralnetworks training are made due to the functionality of thetool

References

[1] P M M Foussier From Product Description to Cost A PracticalApproach The Parametric Approach vol 1 Springer ScienceBusiness Media 2006

[2] A Layer E T Brinke F Van Houten H Kals and S HaasisldquoRecent and future trends in cost estimationrdquo InternationalJournal of Computer IntegratedManufacturing vol 15 no 6 pp499ndash510 2002

[3] S M Trost and G D Oberlender ldquoPredicting accuracy of earlycost estimates using factor analysis andmultivariate regressionrdquoJournal of Construction Engineering and Management vol 129no 2 pp 198ndash204 2003

[4] D J Lowe M W Emsley and A Harding ldquoPredicting con-struction cost using multiple regression techniquesrdquo Journalof Construction Engineering and Management vol 132 no 7Article ID 002607QCO pp 750ndash758 2006

[5] S Varghese and S Kanchana ldquoParametric Estimation of Con-struction Cost Using Combined Bootstrap And RegressionTechniquerdquo in Proceedings of the International Journal of CivilEngineering and Technology vol 5 pp 201ndash205 2014

[6] Y Shin ldquoApplication of boosting regression trees to preliminarycost estimation in building construction projectsrdquo Computa-tional Intelligence andNeuroscience vol 2015 Article ID 1497022015

[7] N I El Sawalhi ldquoModelling the Parametric ConstructionProject Cost Estimate using Fuzzy Logicrdquo in Proceedings ofthe International Journal of Emerging Technology and AdvancedEngineering vol 2 pp 63ndash636 2012

Complexity 11

[8] I Dikmen M T Birgonul and S Han ldquoUsing fuzzy risk assess-ment to rate cost overrun risk in international constructionprojectsrdquo International Journal of Project Management vol 25no 5 pp 494ndash505 2007

[9] S-H An G-H Kim and K-I Kang ldquoA case-based reasoningcost estimating model using experience by analytic hierarchyprocessrdquo Building and Environment vol 42 no 7 pp 2573ndash2579 2007

[10] SH JiM Park andH S Lee ldquoCase adaptationmethod of case-based reasoning for construction cost estimation in KoreardquoJournal of Construction Engineering and Management vol 138no 1 pp 43ndash52 2011

[11] C M Bishop Neural Networks for Pattern Recognition OxfordUniversity Press New York NY USA 1995

[12] S Haykin Neural networks A Comprehensive FoundationPrentice Hall PTR 1994

[13] S Osowski ldquoSieci neuronowe w ujeciu algorytmicznymrdquoWydawnictwa Naukowo-Techniczne 1996

[14] S Osowski ldquoSieci neuronowe do przetwarzania informacjirdquoOficyna Wydawnicza Politechniki Warszawskiej 2000

[15] R Tadeusiewicz Sieci neuronowe vol 180 AkademickaOficynaWydawnicza Warszawa 1993

[16] CGWilmot andBMei ldquoNeural networkmodeling of highwayconstruction costsrdquo Journal of Construction Engineering andManagement vol 131 no 7 pp 765ndash771 2005

[17] A Alqahtani and A Whyte ldquoArtificial neural networks incor-porating cost significant items towards enhancing estimationfor (life-cycle) costing of construction projectsrdquo AustralasianJournal of Construction Economics and Building vol 13 no 3pp 51ndash64 2013

[18] C L C Roxas and J M C Ongpeng ldquoAn Artificial NeuralNetwork Approach to Structural Cost Estimation of BuildingProjects in the Philippinesrdquo in Proceedings of the DLSU ResearchCongress p 1 Manila 2014

[19] I Elsawy H Hosny and M A Razek ldquoA neural networkmodel for construction projects site overhead cost estimatingin Egyptrdquo International Journal of Computer Science Issues vol8 no 3 pp 273ndash283

[20] T PWilliams ldquoPredicting completed project cost using biddingdatardquo Construction Management and Economics vol 20 no 3pp 225ndash235 2002

[21] HMGunaydin S ZDoganHMGunaydın and S ZDoganldquoA neural network approach for early cost estimation of struc-tural systems of buildingsrdquo in Proceedings of the InternationalJournal of Project Management vol 22 pp 595ndash602 2004

[22] M Juszczyk ldquoThe Use of Artificial Neural Networks for Resi-dential Buildings Conceptual Cost Estimationrdquo in Proceedingsof the 11th International Conference of Numerical Analysisand Applied Mathematics 2013 ICNAAM 2013 pp 1302ndash1306Greece September 2013

[23] M Juszczyk ldquoApplication of committees of neural networks forconceptual cost estimation of residential buildingsrdquo in Proceed-ings of the International Conference on Numerical Analysis andApplied Mathematics 2014 ICNAAM 2014 Greece September2014

[24] M Juszczyk ldquoThe Challenges of Nonparametric Cost Esti-mation of Construction Works with the use of ArtificialIntelligence Toolsrdquo in Proceedings of the Creative ConstructionConference CCC 2017 pp 415ndash422 Croatia June 2017

[25] M Juszczyk A Lesniak and A Lesniak ldquoSite overhead costindex prediction using RBF Neural Networks Transactions onEconomics and Managementrdquo ICEM pp 381ndash386 2016

[26] A Lesniak ldquoThe application of artificial neural networks inindirect cost estimationrdquo in Proceedings of the 11th InternationalConference of Numerical Analysis and Applied Mathematics2013 ICNAAM 2013 pp 1312ndash1315 Greece September 2013

[27] E Pabisek Systemy hybrydowe integrujące MES i SSN w anal-iziewybranych problemowmechaniki konstrukcji imateriałowWydawnictwo Politechniki Krakowskiej Krakow 2008

[28] Z Waszczyszyn ldquoZastosowanie sztucznych sieci neuronowychw inzynierii ladowej XLI Konferencja Naukowa KILiW PAN iKomitetu Nauki PZITBrdquo Krakow-Krynica vol tom 9 pp 251ndash288 1995

Research ArticleAssessment of the Real Estate Market Value inthe European Market by Artificial Neural Networks Application

Jasmina SetkoviT1 Slobodan LakiT1 Marijana Lazarevska2 Miloš CarkoviT 3

Saša VujoševiT1 Jelena CvijoviT4 andMladen GogiT5

1Faculty of Economics University of Montenegro 81000 Podgorica Montenegro2Faculty of Civil Engineering University of Ss Cyril and Methodius 1000 Skopje Macedonia3Erste Bank AD Podgorica 81000 Podgorica Montenegro4Economics Institute 11000 Belgrade Serbia5Faculty of Civil Engineering University of Montenegro 81000 Podgorica Montenegro

Correspondence should be addressed to Milos Zarkovic miloszarkovic87gmailcom

Received 22 August 2017 Accepted 17 December 2017 Published 29 January 2018

Academic Editor Luis Braganca

Copyright copy 2018 Jasmina Cetkovic et al This is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work is properlycited

Using an artificial neural network it is possible with the precision of the input data to show the dependence of the property pricefrom variable inputs It is meant to make a forecast that can be used for different purposes (accounting sales etc) but also forthe feasibility of building objects as the sales price forecast is calculated The aim of the research was to construct a prognosticmodel of the real estate market value in the EU countries depending on the impact of macroeconomic indicators The availableinput data demonstrates that macroeconomic variables influence determination of real estate prices The authors sought to obtaincorrect output data which show prices forecast in the real estate markets of the observed countries

1 Introduction

The difficulty or impossibility of constructing an overallmodel with potential reactions and counterreactions (partic-ipants or agents) stems from the complexity of social eco-nomic and financial systems It is assumed that the method-ology of the neural network with evident complexity of thesystem the usual incomprehensibility and general impracti-cality of the model can help to emulate and encourage theobserved economy or societyThe problem is more obvious ifone tries to control all possible variables and potential resultsin the system as well as to include all their dynamic interac-tions [1] The application of artificial neural networks as wellas econometricmodels is characterized by specific advantagesand disadvantages Nevertheless neural networks have beenimposed as a real alternative to econometric methods that isas a powerful tool for assessment and forecasting for exam-ple in the field of evaluating real estate It is specially empha-sized that it is possible to find estimated values instead ofexact values

Artificial neural networks are relatively new computertools that are widely used in solving many complex realproblems Their attractiveness is a product with good char-acteristics in data processing tolerance of input data errorshigh learning opportunities on examples easy adaptation tochanges and generalization of the methodology for develop-ing successful artificial neural networks starting with concep-tualization projects through design projects to implemen-tation projects [2] The use of artificial neural networks forforecasting has led to a vast increase in research over the pasttwo decades [3]

A generation of various prognostic models is based onthe use of artificial neural networks which are recently beingstudied as a contemporary interdisciplinary field atmany uni-versities which help solve many engineering problems thatcannot be solved using traditional methods Researchedneural networks are used successfully as an analytical tool forrelevant forecasts that greatly improve the quality of decision-making at various levels The common problem of successfulapplications of artificial neural networks in forecasts and


2 Complexity

modeling is related to the lack of necessary datamdashthe dataavailable is ldquonoisyrdquo or incomplete and the circumstances thatthe quantities being modeled are governed by multivariateinterrelationships [4] On the other hand a large numberof studies point to the fact that prognostic models obtainedby the use of artificial neural networks exhibit a satisfactorydegree of accuracy and are particularly useful in situationsfor preperformed numerical or experimental researches [5]The use of these models has been explored and demonstratedreliability in a large number of applications in construction[6ndash11]

The wide scientific and technical use of neural networksin the conditions of increased complexity of themarket whenthey show superiority (effectiveness) in an unstable environ-ment is acceptable to analysts investors and economistsdespite the shortcomings The multidisciplinarity of neuralnetworks and their complexity coverage makes them suitablefor assessing market variables that is underlying indexedand derivative financial instruments Comparing the insta-bility of forecasts obtained using neural networks with theencompassed volatility of the SampP Index futures options andapplying the BAWpricingmodel of options to futures Hamidconcluded that forecasts fromneural networks are better thanthe implied instability forecasts and slightly different from therealized volatility [12]

Models of artificial neural networks provide reasonableaccuracy for many engineering problems that are difficult tosolve by conventional engineering approaches of engineeringtechniques and statistical methods [13 14] The applicationof artificial neural networks in the construction sector is ofgreat importance and usable value Recent literature suggeststhat the methodology of neural networks is largely used tomodel different problems and phenomena in the field ofconstruction [15ndash20]

Artificial neural networks are useful for modeling therelationship between inputs and outputs directly on the basisof observed data As already noted they are capable oflearning generalizing results and responding to highly ex-pressed incompleteness or incompleteness of available data[21] As artificial neural networks have better performancethan multivariate analysis because they are nonlinear theyare able to evaluate subjective information difficult to includein traditional mathematical approaches Furthermore theprominent abilities of neural networks are particularly evi-dent in complex systems such as the real estate system wherein recent years artificial neural networks are widely used tocreate a model for estimating prices in real estate marketsA number of such models have been published in scientificand professional literatureThus in the last decade of the 20thcentury Borst defined a number of variables for the design ofa model based on artificial neural networks to appraise realestate inNewYork State demonstrating that themodel is ableto predict the real estate price with 90 accuracy [22]

2 Materials and Methods

The problem of assessing the market value of real estate inconstruction has always been current from the angles of boththe real estate buyers and sellersinvestors and in particular

for future investments Therefore in theory and practicethere are various methods for determining the market valueof real estate Given that the market value of a property isinfluenced by a large number of factors the process of estima-ting the real estate market is quite complex and is alwayscurrent again A negative assessment practice that only con-firms the agreed real estate price is lighter but less preciseand can lead to deviations of estimated value from the realmarket value of real estate [23] As part of traditionalmethodsof estimation derive from the impact of objective factors onthe real estate price neglecting the undeniable influence ofsubjective factors that have a significant impact on the price ofreal estate generalization of the application of these methodsis not possible because the real estate market in differentcountries is influenced by various objective and subjectivefactors The hedonistic real estate assessment method basedon a multiple regression analysis though often used to testnew estimation methods is burdened by initial assumptionsand is not sufficiently rational to evaluate [23 24]

With the development of new computer and mathemat-ical modeling methods the trend of developing new ap-proaches for real estatemarket assessment is notable Namelythe use of artificial neural networks has proved to be justifiedin the development of unconventional methods for assessingreal estate market value which enable a more objective andmore accurate estimate of real estate in themarket As the val-uation process is always a problem in free market economiesmarket participants usually do not have complete andaccurate pricing information which is why they are consid-ering a variety of factors and different relationships betweenthem In the real estate market there is usually a morepronounced lack of accurate information compared to infor-mation about another commodity because data is usually notavailable in a consistent format Analysis and interpretation ofgeneral trends are hampered by the sparse variety of charac-teristicsproperties of these commodities usually related toa particular location [25] Therefore a large number of au-thors elaborated other approaches as an alternative to theconventional method of assessment and developed newmodels for assessing the market value of real estate that arecapable and usable for similar purposes [26ndash28]

Developed models of artificial neural networks haveshown that residential property markets are under the influ-ence of different economic and financial environments Theaim of developing such models is to indicate inter aliawhether the economic and financial situation of the observedcountries reflects the general economic situation on the realestate market The results showed that the economic andfinancial crisis in these countries had different impacts onreal estate prices [29] The methodology based on roughset theory and on artificial neural networks proved to besuitable for the convergence of residential property pricesindex [29 30] During the last decade the literature analyzedthe impact of property price volatility on the economy ingeneral unemployment consumer confidence in govern-ment banking practices and social costs On the other handthe macroeconomic situations such as the business cycleemployment rates income growth interest rates inflationrates loan supply returns on real estate investments and

Complexity 3

other factors such as population growth have had a signif-icant impact on housing prices

Contrary to the dominant application of multiple regres-sion models involving human estimations the use of arti-ficial neural networks the model of artificial intelligenceallows the hidden nonlinear connections between the mod-eled variables to be uncovered In the context of neuralnetworks a special place was occupied by the so-called back-propagation models These neural networks contain series ofsimple interconnected neurons (or nodes) between the inputand output vectors Pi-Ying used a backpropagation neuralnetwork as a tool for constructing a housing price model fora selected city [31] The paper investigates the importance ofthe application of neural network technologies in the realestate appraisal problem Starting from the consequences ofthe nature of the neural network model Pi-Ying concludesthat the model creates a larger prediction error comparedto multiple regression analysis However by estimating alarge number of real estate properties Peterson and Flanaganfound that artificial neural networks compared to linearhedonic pricing models create significantly minor errors indollar pricing and have greater out-of-sample precision andthere is a better extrapolation in a more volatile environment[32]

According to Limsombunchai (2004) the hedonistic costmodel and the artificial neural network emphasized thatthe hedonistic technique is generally unrealistic in dealingwith the housing market in any geographic area as a singleunit Compared to neural network models this model showspoorer results in the out-of-sample prediction [33] Marketimperfections complemented by bid rigidity and heterogene-ity of quality contribute to the deviation of real estate pricesfrom the fundamental value Consequently under sustainabledeviation conditions the problem of adverse selection isspurred and in the periods of financial liberalization and theboom-bust cycle the consequence (distortion) is moral haz-ard [30] The empirical and operational framework suggeststhat the residential property market directly determined themagnitude of the economic growth

European economic growth is supported by the expandedECB stimulus (quantitative easing) which has increasedliquidity confidence and domestic consumption by creatinga fundament that impacts the real estate market Growththat is the boom of house prices in Europe shows continuitylately This can be tracked over the momentum which isincreasing if the property market grows faster this yearcompared to the previous (or fall below) Therefore this isa ldquochange in changerdquo measure which shows that most of thehousing market is slowing down although the boom contin-ues strongly in Europe [34] The factor that has affected theEuropean housingmarket is GDP growth In the beginning ofthe previous decade the correlation between lagging in GDPgrowth and house prices in the EU was 81 According tothese analyzes the expected sluggish growth of the economyshould limit the growth of apartments prices in the comingyears

Interest rates as another economic indicator of the realestate market depend on monetary policy in the ECB andother central banks in the EU Low interest rates are the

result of expansive monetary policy which stimulates the realestatemarketThere is a correlation of residential lending andhouse prices The rise in house prices is a consequence of thegrowth of residential lending and as a result of price risesthe ratio of resident debt to household disposable incomeis changing This indebtedness of households along with thehousehold debt capacity becomes a determinant of the rise inhouse pricesThere are two consequences of cheap residentialhousing in the European Union (2015) the rapid rise inproperty prices in certain markets and the great difference intransaction prices between cities and countries [35]

Housing is not only an important segment of householdwealth but also a key sector of the real economy The declinein housing construction can be reflected negatively (directlyor indirectly) on financial stability and the real economy Inaddition the role of the loan is significant in the rise andrecession of housing Financial and macroeconomic stabilityare significantly affected by the development of the residentialreal estate sectorThemain responsibility of macroprudentialauthorities is to analyze the vulnerabilities of this marketIn the EU European Systemic Risk Board has the mandateto implement ldquomacroprudential oversight of the financialsystem within EU in order to contribute to the preventionor mitigation of systemic riskrdquoThe ESRB identifies countriesin the EU that have medium-term sensitivities that can bethe cause of systemic risk and result in serious negativeconsequences Horizontal analysis based on key indicatorsrisk analysis and analysis of structural and institutionalfactors (vertical) are implemented [36]

In this paper a model for estimating real estate prices in27 European countries has been defined This research paperindicates the authorrsquos assumption was aided by a prognosticmodel using artificial neural networks with the availablefactors influencing the real estate price formation Thesefactors can obtain accurate real estate prices in the Europeanmarket which can have a significant value in the decision-making process of buying real estate in the European market

Otherwise in literature there are various approaches tothe division of factors that more or less influence the forma-tion of prices in the real estate markets One of the usualdivisions is onmacroenvironment factors (eg exchange rateemployment GDP loan interest rate and geofactors) andmicroenvironment factors (mostly related to constructionenvironment) [37] In addition to this division there is adivision of factors into rational ones related to the realprice of real estate and irrational ones that reflects theexpectation of consumers [38] At the same time certainanalysis has highlighted the fundamental real estate factorsin any country in relation to loan availability housing supplyand dimet ratio interest rate decrease changes in housingmarket participantsrsquo expectations administrative restrictionsof supply and so forth [39]

For the purpose of developing a prognostic model for theEuropean real estate market based on artificial neural net-works the authors suggestedmacroeconomic factors affectedthe real estatemarket Training of the artificial neural networkwas carried out by defining 11 inputs and one output ofthe network (house price) The output variablesmdashreal estateprices for 27 EU countriesmdashare provided on the basis of

4 Complexity

available empirical data [36] and certain authorsrsquo calculationsTable 1 shows the basic inputs into the model and theirdefinition and basic characteristics Precollection and dataanalysis preparation of data for defining the model andultimately the production of the model were performed

The basic characteristics of the input and output param-eters used for the training network represent 253 sets of dataof which 80 are selected as a network training set and 20as a validation set All data is downloaded from the above-mentioned databases After training the network was con-trolled on 11 sets of data on which the network was nottrained

The market value of real estate is subject to time changesand is determined at a certain date so this prognostic modelis time-dependent

Creating an artificial neural network model trained tosolve this problem consisted in defining network architecturewith 11 variable inputs and one output For network traininga two-layer neural network with two hidden layers of 15neurons in the hidden layer was adopted

Network training was conducted on a nonrecurrentnetwork while network training was carried out using animproved Backpropagation algorithm by periodically passingdata from a training session through a neural network Thevalues obtained were compared with the actual outputs andif the difference was made correction of weight coefficientswasmade Correction of weighing coefficients of the networkwas done by the rule of gradient downhill

119908(119897)new119894119895 = 119908(119897)old119894119895 minus 120578

120597120576119896

120597119908(119897)119894119895 (1)

As an activation function a logistic sigmoidal function wasused

119891 (119909) =1

1 + 119890minus119909 (2)

Improving theBackpropagation algorithmmeant introducinga moment so that the weight change in the period 119905 woulddepend on the change in the previous period

Δ119908(119897)119894119895 (119905) = minus120578120597120576119896

120597119908(119897)119894119895+ 120572Δ119908(119897)119894119895 (119905 minus 1) 0 lt 119905 lt 1 (3)

Network training was selected for a validation set of 20 ofdata Training went to the moment of an error in the valida-tion set so that the trained network showed good forecast per-formances

The neural network is trained within the MS Excelprogram Minor deviations from the training session arenoted Based on the network initiation with the input datafrom the range of data used in training it is possible to drawup prognostic models of the dependence of the output fromany input

Control forecast was performed on 11 test datasets onwhich the network was not trained

Market price FranceForecast price FranceMarket price Czech RepublicForecast price Czech RepublicMarket price LithuaniaForecast price Lithuania

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2005

Year

100000

300000

500000

700000

900000

1100000

Pric

e

Figure 1 Deviation of the real price from the price forecast


In general our research has shown that prognostic modelsbased on the use of artificial neural networks have a satisfac-tory degree of precision Namely the prognostic model forestimating the price of the real estate in the EU market donefor the purposes of this survey is an average deviation of realprice from a price forecast of up to 14 Figure 1 shows thisdeviation for 3 selected countries of different level of devel-opment

On the basis of the prognostic model designed on theuse of artificial neural networks real estate prices in the EUcountries were estimated starting with the macroeconomicvariables that were originally assumed to have a significantimpact on the output variableThe analysis takes into accountthe impact of the change of a certain factor relevant to realestate prices in the EU market with other unchanged realvariables

Figures 2 3 and 4 show changes in forecasted real estateprices in countries with different levels of development (highmedium and low) under the influence of GDP growth ratesIn the countries surveyed regardless of the degree of devel-opment there is a trend of growth of forecasted real estateprices with an increase in the GDP growth rateTheoreticallyGDP growth rate can significantly increase the real estateprice by increased consumption in the conditions of largehousing infrastructure projects that are affecting employmentgrowth Mortgage payments are more acceptable so realestate demand rises and ultimately increases real estate pricesSome research deals with the relationship between changesin real estate prices and changes in real GDP or whether theinterdependence between these two variables is statisticallysignificantThe regression analysis method has indicated that

Complexity 5

Table1Ba

sicinpu

tsinto

them

odel

theird

efinitio

nandbasic

characteris

tics

Varia

ble

Varia

bled

efinitio

n(according

tothes

ources

givenbelowthistable)

Minim

umMaxim

umMean

Standard

deviation

Unito

fmeasure

GDPlowast

GDP(grossdo

mestic

prod

uct)reflectsthe

totalvalue

ofallgoo

dsandservices

prod

uced

lessthe

valueo

fgoo

dsandservices

used

forintermediateconsum

ptionin

theirp

rodu

ction

5142

3032820

47776356

7206493

5mileuro

BDPperc

apitalowast

GDPperc

apita

atmarketp

rices

iscalculated

asther

atio

ofGDPatmarketp

rices

tothea

verage

popu

latio

nof

aspecific

year

4200

84400

2382963

1575946

euro

RealGDPgrow

thratelowast

Thec

alculationof

thea

nnualgrowth

rateof

GDPvolumeisintendedto

allowcomparis

onso

fthe

dynamicso

fecono

micdevelopm

entb

othover

timea

ndbetweenecon

omieso

fdifferentsizesminus148

119

163

384

Inequalityof

income

distr

ibutionlowast

Ther

atio

oftotalincom

ereceivedby

the2

0of

thep

opulationwith

theh

ighestincome(top

quintile)to

thatreceived

bythe2

0of

thep

opulationwith

thelow

estincom

e(lowestq

uintile)

32

83

483

116

-

Totalu

nemploymentratelowast

Unemploymentrates

representu

nemployed

person

sasa

percentage

ofthelabou

rforce

340

2750

907

432

Averagea

nnualn

etearningslowast

Netearnings

arec

alculated

from

grosse

arning

sbydedu

ctingthee

mployeersquossocialsecurity

contrib

utions

andincometaxes

andadding

family

allowancesinthec

aseo

fhou

seho

ldsw

ithchild

ren

155077

3849018

1717581

1044515

euro

FDIlowastlowast

Foreigndirectinvestmentreferstodirectinvestmentequ

ityflo

wsinther

eportin

gecon

omyItis

thes

umof

equitycapitalreinvestm

ento

fearning

sandotherc

apita

lminus29679

734010

27948

67501

mil$

HICP-inflatio

nratelowast

Harmon

isedIndiceso

fCon

sumer

Prices

(HICPs)a

redesig

nedforinternatio

nalcom

paris

onso

fconsum

erpriceinfl

ation

minus16

153

238

220

VAT(

)lowastlowastlowast

Thev

alue

addedtaxabbreviatedas

VATin

theE

urop

eanUnion

(EU)isa

generalbroadlybased

consum

ptiontaxassessed

onthev

alue

addedto

good

sand

services

1527

205

257

Taxeso

nprop

ertyas

of

GDPlowastlowastlowastlowast

Taxon

prop

ertyisdefin

edas

recurrentand

nonrecurrent

taxeso

ntheu

seownershipor

transfe

rof

prop

ertyTh

eseinclude

taxeso

nim

movableprop

ertyor

netw

ealth

taxes

onthec

hangeo

fow

nershipof

prop

ertythroug

hinheritance

orgift

andtaxeso

nfin

ancialandcapitaltransactio

ns

0283

5387

1463

106

Taxeso

nprop

ertyas

of

totaltaxationlowastlowastlowastlowast

0844

14907

4013

274

SourceslowastEu

rosta

tlowastlowastWorld

Bank

lowastlowastlowastEC

BandlowastlowastlowastlowastOEC

DandEu

rosta

t

6 Complexity

minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3

Real GDP growth rate

Year2015201420132012

2011201020092008

200720062005

1200000

1400000

1600000

1800000

2000000

2200000

2400000

Fore

cast

pric

e

Figure 2The impact of the real GDP growth on the real estate priceforecast in the UK

minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3


000

50000

100000

150000

200000

250000

300000

350000

400000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005

Figure 3 The impact of the real GDP growth rate on the real estateprice forecast in Czech Rep

there is a relationship and dependence while the correlationsuggests a possible causal interconnection between thesevariables [40] In OLS (ordinary least squares) models aswell as in models based on the application of artificial neuralnetworks the GDP growth rate is taken as one of the keymacroeconomic variables that affect the price of real estatewith varying degrees of significance in different countries[29]

Figures 5 6 and 7 show changes in the real estate pricesforecast under the influence of the HICP On the figures itcan be seen that regardless of the degree of developmentof the country with the increase in HICP there is a risein the forecasted price of real estate Some studies such asthose from Burinsena Rudzkiene and Venckauskaite onthe example of Lithuania have shown that the HICP as

minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3


100000

120000

140000

160000

180000

200000

220000

240000

260000

Fore

cast

pric

eYear

2014201320122011

201020092008

200720062005

Figure 4 The impact of the real GDP growth rate on the real estateprice forecast in Lithuania

an indicator of the average annual inflation rate is one ofthe main factors of the real estate market [39] Also someresearch conducted for highly developed countries (eg Nor-way case) have shown that a long-term increase in HICP hasbeen accompanied by a rather sharp rise in real estate prices[41]

Figures 8 9 and 10 show changes in the real estate pricesforecast under the influence of the unemployment rate incountries with different levels of development Figures for thesurveyed countries indicate an apparent decline in the realestate prices forecast with an increase in the unemploymentrate In theory the dominant views are that low unem-ployment leads to an income rise and affects the increasein consumer confidence that manifests itself on the realestate market encouraging real estate prices growth and viceversa Earlier empirical research on the impact of economicvariables on property price dynamics has shown that growthin unemployment reduces real estate prices [42] as ourprognostic model also pointed out Certain recent empiricalstudies point to the undeniable impact of the unemploymentrate as a macroeconomic factor on real estate prices Theimpact of the unemployment rate on real estate prices differsfrom country to country One of these studies has shown thatthe price of real estate in France Greece Norway and Polandis statistically significantly associated with unemployment[43] Also research related to Ireland proves that at low levelsof unemployment real estate prices tend to grow [44]

Figures 11 12 and 13 show changes in real estate pricesforecast in countries with different levels of development

Complexity 7

minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

000

200000

400000

600000

800000

1000000

1200000

1400000

1600000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005

Figure 5 The impact of the HICP-inflation rate on the real estateprice forecast in Germany

minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

150000

170000

190000

210000

230000

250000

270000

290000

310000

Fore

cast

pric

e

Year2015201420132012

201120102009

200820072006

Figure 6 The impact of the HICP-inflation rate on the real estateprice forecast in Slovakia

(high medium and low developed) under the influenceof average annual net earnings The growth of averageannual net earnings is followed by the growth of the out-put variablemdashreal estate prices in the observed countriesEmpirical researches have shownduring the previous decadesthat the growth rate of income (earnings) is an essentialdeterminant of the growth of real estate pricesThus in highlydeveloped countries income growth (earnings) at lowerinterest rates and slower lending conditions is recognized

minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

130000150000170000190000210000230000250000270000290000

Fore

cast

pric

e

Year2014201320122011

201020092008

200720062005

Figure 7 The impact of the HICP-inflation rate on the real estateprice forecast in Lithuania

340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54

Total unemployment rate

200000400000600000800000

100000012000001400000160000018000002000000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005

Figure 8 The impact of the total unemployment rate on the realestate price forecast in France

340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54


150000

200000

250000

300000

350000

400000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005

Figure 9 The impact of the total unemployment rate on the realestate price forecast in Czech Rep

8 Complexity

340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54


Year201420132012

201120102009

200820072006

100000

120000

140000

160000

180000

200000

220000

240000

Fore

cast

pric

e

Figure 10 The impact of the total unemployment rate on the realestate price forecast in Bulgaria

000

100000

200000

300000

400000

500000

600000

700000

800000

Fore

cast

pric

e

329

792

631

142

28

293

053

127

468

33

109

355

612

772

54

146

095

116

446

49

182

834

620

120

44

219

574

123

794

38

256

313

6

348

162

336

653

21

726

161

542

464

358

766

909

859

175

069

Net earningsYear

2015201420132012

20072006

2011201020092008

2005

Figure 11The impact of net earnings on the real estate price forecastin Germany

as a key factor in the rapid rise of the real estate prices[45] Certain models such as the P-W model indicate thathouse prices are functions of the cyclical unemployment rateincome demographics costs of financing housing purchasesand costs of construction materials [46]

4 Conclusion

The economic and social importance of the real estatemarket corresponds to overall economic development butthe housing sector can also be the cause of vulnerability and

Year2015201420132012

201120102009

200820072006

000

200000

400000

600000

800000

1000000

1200000

Fore

cast

pric

e

340

814

432

611

86

311

422

829

672

70

282

031

2

105

681

712

037

75

135

073

314

976

91

164

464

917

916

07

193

856

420

855

22

223

248

023

794

38

252

639

626

733

54

355

510

237

020

60

762

901

615

943

468

985

322

027

909

859

175

069

Net earnings

Figure 12 The impact of the net earnings on the real estate priceforecast in Slovakia

175

069

322

027

468

985

615

943

762

901

909

859

105

681

712

037

75

135

073

314

976

91

164

464

917

916

07

193

856

420

855

22

223

248

023

794

38

252

639

626

733

54

282

031

229

672

70

311

422

832

611

86

340

814

435

551

02

370

206

0

Net earningsYear

2014201320122011

201020092008

200720062005

000

500000

1000000

1500000

2000000

2500000

Fore

cast

pric

e

Figure 13 The impact of the net earnings on the real estate priceforecast in Lithuania

crisis Therefore the problem of estimating the value of realestate is always current and complex due to the influence of alarge number of variables that are recognized in the literatureas usual as macroeconomic construction and other factorsContrary to conventional real estate estimationmethods newapproaches have been developed to evaluate prices in realestate markets In addition to the hedonic method in thelast decades models of artificial neural networks that providemore objective and accurate estimates have been developedThis paper presents a prognostic model of real estate marketprices for EU countries based on artificial neural networks

For a rough and fast estimation of house prices with areliability of about 85 it is possible to use a trained neuralnetwork We consider the obtained level of reliability as very

Complexity 9

high it is about modeling the sociotechnical system Moreaccurate pricing and reliable information could be obtainedif a larger set of input parameters was included

It is shown that the neural network can model nonlinearbehavior of input variables and generalize real estate pricesdata for random inputs in the network training range Themodel shows a satisfactory degree of forecasted precisionwhich guarantees the possibility of its applicability



References

[1] Y Shachmurove ldquoBusiness application of emulative neural net-worksrdquo International Journal of Business vol 10 no 4 2005

[2] I A Basheer and M Hajmeer ldquoArtificial neural networks fun-damentals computing design and applicationrdquo Journal ofMicrobiological Methods vol 43 no 1 pp 3ndash31 2000

[3] G Zhang B E Patuwo and M Y Hu ldquoForecasting with arti-ficial neural networks the state of the artrdquo International Journalof Forecasting vol 14 no 1 pp 35ndash62 1998

[4] AWCOreta andKKawashima ldquoNeural networkmodeling ofconfined compressive strength and strain of circular concretecolumnsrdquo Journal of Structural Engineering vol 129 no 4 pp554ndash561 2003

[5] M Lazarevska M Knezevic M Cvetkovska and A Trombeva-Gavriloska ldquoApplication of artificial neural networks in civilengineeringrdquo Technical Gazette vol 21 no 6 pp 1353ndash13592014

[6] M Lazarevska M Knezevia M Cvetkovska N Ivanisevic TSamardzioska and A Trombeva-Gavriloska ldquoFire-resistanceprognostic model for reinforced concrete columnsrdquo Gradjev-inar vol 64 no 7 pp 565ndash571 2012

[7] D Bojovic D Jevtic and M Knezevic ldquoApplication of neuralnetworks in determination of compressive strength of concreterdquoRomanian Journal of Materials vol 42 no 1 pp 16ndash22 2012

[8] B Scepanovic M Knezevic and D Lucic ldquoMethods for deter-mination of ultimate load of eccentrically patch loaded steel I-girdersrdquo Informes de laConstruccion vol 66 no 1 pp 1ndash13 2014

[9] M Knezevic M Lazarevska M Cvetkovska A Trombeva-Gavriloska andMMilanovic ldquoNeural network based approachfor prediction the fire resistance of centrically loaded compositecolumnsrdquo Technical Gazette vol 23 no 5 2016

[10] M Knezevic D Bojovic D Nikolic K Jovanovic and LLoncar ldquoThe effect of entrapped air on concrete compressivestrength- neural network approach and classical researchrdquo inProceedings of the 14th International Symposium of DGKM pp69ndash74 Struga Macedonia 2011

[11] M Lazarevska M Knezevic and M Cvetkovska ldquoApplicationof artificial neural networks for prognostic modeling of fireresistance of reinforced concrete pillarsrdquoAppliedMechanics andMaterials vol 148-149 pp 856ndash861 2011

[12] S A Hamid ldquoPrimer on using neural networks for forecastingmarket variablesrdquo in Proceedings of the Conference at School ofBusiness Southern New Hampshire University 2004

[13] I Flood ldquoSimulating the construction process using neuralnetworksrdquo in Proceedings of the 7th International Symposium on

Automation and Robotics in Construction pp 374ndash382 BristolUK June 1990

[14] D S Jeng D H Cha andM Blumenstein ldquoApplication of neu-ral networks in civil engineering problemsrdquo in Proceedings ofthe International Conference on Advances in the Internet Pro-cessing Systems and Interdisciplinary Research 2003

[15] A Mukherjee and J M Deshpande ldquoModeling initial designprocess using artificial neural networksrdquo Journal of Computingin Civil Engineering vol 9 no 3 pp 194ndash200 1995

[16] M A Shahin H R Maier and M B Jaksa ldquoPredicting settle-ment of shallow foundations using neural networksrdquo Journal ofGeotechnical and Geoenvironmental Engineering vol 128 no 9pp 785ndash793 2002

[17] A Sanad and M P Saka ldquoPrediction of ultimate shear strengthof reinforced-concrete deep beams using neural networksrdquoJournal of Structural Engineering vol 127 no 7 pp 818ndash8282001

[18] R J Abrahart and LM See ldquoNeural networkmodelling of non-linear hydrological relationshipsrdquo Hydrology and Earth SystemSciences vol 11 no 5 pp 1563ndash1579 2007

[19] A M Elazouni I A Nosair Y A Mohieldin and A G Mo-hamed ldquoEstimating resource requirements at conceptual designstage using neural networksrdquo Journal of Computing in CivilEngineering vol 11 no 4 pp 217ndash223 1997

[20] J Xie J-F Qiu W Li and J-W Wang ldquoThe application ofneural network model in earthquake prediction in East ChinardquoAdvances in Intelligent and Soft Computing vol 106 pp 79ndash842011

[21] J Shaw ldquoNeural network resource guiderdquo AI Expert vol 8 no2 pp 48ndash54 1992

[22] R A Borst ldquoArtificial neural networks the next modelingcali-bration technology for the assessment communityrdquo PropertyTax Journal vol 10 no 1 pp 69ndash94 1991

[23] M Mattos P Simoes E Zancam N Ferreira and C CechinelldquoA neurofuzzy approach for property value predictionrdquo in Pro-ceedings of the 35th Conferencia Latinoamericana de Informatica(CLEI rsquo08) 2008

[24] D Fischer and P P Lai ldquoArtificial neural networks and com-puter assisted mass appraisalrdquo in Proceedings of the 12th AnnualConference of the Pacific Rim Real Estate Society (PRRES rsquo06)Auckland New Zealand January 2006

[25] C Bagnoli and H C Smith ldquoThe theory of fuzzi logic and itsapplication to real estate valuationrdquo The Journal of Real EstateResearch vol 16 no 2 1998

[26] H Kusan O Autekin and I Ozdemir ldquoThe use of fuzzi logic inpredicting house selling pricerdquo Expert System Application vol37 no 3 2010

[27] S Yalpir and G Ozkan ldquoFuzzy logic methodology and multipleregressions for residential real-estates valuation in urban areasrdquoScientific Research and Essays vol 6 no 12 pp 2431ndash2436 2011

[28] J Zurada ldquoNon-conventional approaches to property valueassessmentrdquo Journal of Applied Business Research vol 22 no 32006

[29] MRenigier-Bilozor andRWisniewski ldquoThe impact ofmacroe-conomic factors on residential property prices indices inEuroperdquo XLI Incontro di Studio del CeSET pp 149ndash166 2013

[30] R Wisniewski ldquoModeling of residential property prices indexusing committees of artificial neural networks for PIGS theEuropean-G8 and Polandrdquo Argumenta Oeconomica vol 1 no38 2017

10 Complexity

[31] L Pi-Ying ldquoAnalysis of the mass appraisal modelrdquo in Proceed-ings of the 23rd Pan Pacific Congress of Appraisers Valuers andCounselors San Francisco Calif USA 2006

[32] S Peterson and A B Flanagan ldquoNeural network hedonic pric-ing models in mass real estate appraisalrdquo Journal of Real EstateResearch vol 31 no 2 pp 147ndash164 2009

[33] VLimsombunchai lsquolsquoHouse price prediction hedonic pricemod-el vs artificial neural networkrdquo in Proceedings of the NZARESConference Blenheim New Zealand June 2004

[34] European Systemic Risk Board ldquoVulnerabilities in the EU resi-dential real estate sectorrdquo European System of Financial Super-vision 2016httpswwwesrbeuropaeupubpdfreports161128vulnerabilities eu residential real estate sectorenpdf

[35] Global Property Guide ldquoResidential property investment re-searchrdquo httpwwwglobalpropertyguidecom

[36] Deloitte ldquoProperty Indexmdashoverview of European residentialmarketsrdquo Deloitte Czech Republic 2016 httpswww2deloittecomcontentdamDeloitteczDocumentssurveyPropertyIndex 2016 ENpdf

[37] S Vanichvatana ldquoThailand real estate market cycles case studyof 1997 economic crisisrdquo GH Bank Housing Journal vol 1 no 1pp 38ndash47 2007

[38] V Rudzkiene and V Azbainis ldquoNekilnojamo turto rinkos evo-liucijos pokyciai pereinamosios ekonomikos sąlygomisrdquo in Pro-ceedings of the Business Management and Education ConferenceProgram Vilnius Gediminas Technical University Novemeber2010

[39] M Burinskiene V Rudzkiene and J Venckauskaite ldquoModels offactors influencing the real estate pricerdquo in Proceedings of the 8thInternational Conference on Environmental Engineering VilniusGediminas Technical University Vilnius Lithuania May 2001

[40] R M Valadez ldquoThe housing bubble and the US GDP a corre-lation perspectiverdquo Journal of Case Research in Business andEconomics vol 3 Article ID 10490 2010

[41] I Johansen and R Nygaard ldquoOwner-occupied housing in theNorwegian HICPrdquo Tech Rep 200918 Statistics Norway OsloNorway 2009

[42] J Clapp and C Giacotto ldquoThe influence of economic variableson house price dynamicsrdquo Journal of Urban Ecnomics vol 36pp 116ndash183 1993

[43] B Grum and D K Govekarb ldquoInfluence of macroeconomicfactors on prices of real estate in various cultural environmentscase of Slovenia Greece France Poland and Norwayrdquo ProcediaEconomics and Finance vol 39 pp 597ndash604 2016

[44] M Mernagh ldquoHouse prices and unemployment a story oflinked fortunesrdquo Significance in Economics amp Business RoyalStatistical Society and American Statistical Association 2014

[45] K Sommer P Sullivan and R Verbrugge ldquoRun-up in the houseprice-rent ratio howmuch can be explained by fundamentalsrdquoBureau of Labor Statistics Working paper 441 2011

[46] J Peek and J A Wilcox ldquoThe baby boom ldquopent-uprdquo demandand future house pricesrdquo Journal of Housing Economics vol 1no 4 pp 347ndash367 1991

Research ArticleDevelopment of ANN Model for Wind Speed Prediction asa Support for Early Warning System

IvanMaroviT1 Ivana Sušanj2 and Nevenka ODaniT2

1Department of Construction Management and Technology Faculty of Civil Engineering University of Rijeka51000 Rijeka Croatia2Department of Hydraulic Engineering and Geotechnical Engineering Faculty of Civil Engineering University of Rijeka51000 Rijeka Croatia

Correspondence should be addressed to Ivana Susanj isusanjunirihr

Received 28 September 2017 Accepted 28 November 2017 Published 20 December 2017

Academic Editor Milos Knezevic

Copyright copy 2017 Ivan Marovic et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

The impact of natural disasters increases every year with more casualties and damage to property and the environment Thereforeit is important to prevent consequences by implementation of the early warning system (EWS) in order to announce the possibilityof the harmful phenomena occurrence In this paper focus is placed on the implementation of the EWS on the micro location inorder to announce possible harmful phenomena occurrence caused by wind In order to predict such phenomena (wind speed) anartificial neural network (ANN) prediction model is developed The model is developed on the basis of the input data obtained bylocal meteorological station on the University of Rijeka campus area in the Republic of Croatia The prediction model is validatedand evaluated by visual and common calculation approaches after which it was found that it is possible to perform very good windspeed prediction for time steps Δ119905 = 1 h Δ119905 = 3 h and Δ119905 = 8 h The developed model is implemented in the EWS as a decisionsupport for improvement of the existing ldquoprocedure plan in a case of the emergency caused by stormy wind or hurricane snow andoccurrence of the ice on the University of Rijeka campusrdquo

1 Introduction

Today we are witnessing a high number of various naturalharmful meteorological and hydrological phenomena whichincrease every year with a greater number of casualties anddamage to property and the environmentThe impact of thesedisasters quickly propagates around the world irrespectiveof who or where we are [1] To cope with such a growingnumber of professionals and volunteers strive to help [2]by gaining knowledge and taking actions toward hazardrisk vulnerability resilience and early warning systems butsuch has not resulted in changing the practices of disastermanagement [3 4] Additionally some natural hazards havecontinuous impact on a smaller scale that is location Theirrate of occurrence and the impact they make around bringimportance of implementing the same disaster managementprinciples as is on bigger scale

In order to cope with such challenges Quarantelli (1988)[5] argued that the preparation for and the response to

disasters require well aligned communication and informa-tion flows decision processes and coordination structuresSuch become basic elements of todayrsquos early warning systems(EWSs) According to the UnitedNations International Strat-egy for Disaster Reduction (UNISDR 2006) [6] a completeand effective EWS includes four related elements (i) riskknowledge (ii) a monitoring and warning system service(iii) dissemination and communication and (iv) responsecapability

Today the implementation of artificial neural networks(ANN) in order to develop prediction models becomes veryinteresting research field The usage of ANN in field ofhydrology and meteorology is relatively new since the firstapplication was noted in the early nineties The first imple-mentation of ANN in order to predict wind speed is noted byLin et al (1996) [7] Mohandes et al (1998) [8] and Alexiadiset al (1998) [9] Also in order to predict wind speed recentlynumerous examples of ANN implementation are notedFonte et al (2005) [10] in their paper introduced a prediction


2 Complexity

Envi

ronm

ent Tactical

management level


Strategicmanagement

level

Database Modelbase

Dialog

Users

Figure 1 Structure of the decision support system for early warning management system

of the average hourly wind speed and Monfared et al (2009)[11] in their paper introduced a new method based on ANNfor better wind speed prediction performance presentedPourmousavi Kani and Ardehali (2011) [12] introduced a veryshort-term wind speed prediction in their paper and Zhou etal (2017) [13] presented a usage of ANN in order to establishthe wind turbine fault early warning and diagnosis modelAccording to the aforementioned implementations of theANN development of such a prediction model in order toestablish an EWS on the micro locations that are affected bythe harmful effects of wind is not found and it needs to bebetter researched

The goal of this research is to design and develop anANN model in order to achieve a successful prediction ofthe wind speed on micro locations based on data fromlocal meteorological stations The final aim and objective ofthe research is to implement a developed ANN wind speedpredictionmodel in order to serve as decision support tool inan EWS

This paper is organized as follows Section 2 providesa research background of decision support as well as themethodology for development ANN prediction models asa tool for supporting decisions in EWS In Section 3 theresults of the proposed model are shown and discussedwith implementation in EWS on the micro location ofUniversity of Rijeka campus area Finally the conclusion andrecommendations are presented in Section 4

2 Methodology

Depending on the need of the business different kinds ofinformation systems are developed for different purposesIn general different kinds of data and information are

suitable for decision-making in different levels of organiza-tion hierarchy and require different information systems tobe placed such as (i) Transaction Process Systems (TPS)(ii) Management Information Systems (MIS) (iii) DecisionSupport Systems (DSS) and (iv) Executive Support Systems(ESS) [16] At the same time each information system cannotfulfill complete information needs of each level Thereforeoverlapping of the systems is needed in order to preserve bothinformation flows (from lower to upper level) and actions(from upper to lower level)

A structure of the proposed early warning managementsystem (Figure 1) is based on Marovicrsquos previous research[17] where the ldquothree decision levelsrdquo concept for urbaninfrastructure [18 19] and spatial [17] management is pro-posed The modular concept is based on DSS basic structure[20] (i) database (ii) model base and (iii) dialog moduleInteractions between modules are realized throughout thedecision-making process at all management levels as theyserve as meeting points of adequate models and data

The first management level supports decision-makersat lowest operational management level Beside its generalfunction of supporting decision-making processes at oper-ational level it is a meeting point of data and informationAdditionally it provides information flows toward higherdecision levels It is a procedural level where problems arewell defined and structured The second management leveldeals with less defined and semistructured problems On thislevel tactical decisions are delivered and it is a place whereinformation basis and solutions are created Based on appliedmodels from model base (eg ANN) it gives alternativesand a basis for future decisions on strategic managementlevel which deals with even less defined and unstructuredproblems At the thirdmanagement level based on the expert

Complexity 3

deliverables from the tactical level a future development ofthe system is carried out Strategies are formed and they serveas frameworks for lower decision and management levels

Outside factors from the environment greatly influencedecision support system as is shown in Figure 1 on bothdecision-making and whole management processes Suchstructure is found to be adequate for various urban manage-ment systems and its structure easily supports all phases ofthe decision-making in early warning systems

In addition to the previously defined four elements EWSsare specific to the context for which they are implementedGlantz (2003) [21] described general principles that oneshould bear in mind (i) continuity in operations (ii) timelywarnings (iii) transparency (iv) integration (v) humancapacity (vi) flexibility (vii) catalysts and (viii) apoliticalIt is important to take into account current and localinformation [22] as well as knowledge from past events(stored in a database) or grown structures [23] In orderto be efficient in emergencies EWSs need to be relevantand user-oriented and allow interactions between all toolsdecision-makers experts and stakeholders [24ndash26] It canbe seen as an information system which deals with varioustypes of problems (from structured to unstructured) Theharmful event prediction model based on ANN is in generaldeveloped under the monitoring and warning system serviceand becomes a valuable tool in the DSSrsquos model base

The development of ANN prediction model requires anumber of technologies and areas of expertise It is necessaryto have an adequate set of data (from monitoring andorcollection of existing historical data) on the potential hazardarea real-time and remote monitoring of trigger factorsOn such a collection of data the data analysis is madewhich serves as a starting point on the prediction modeldevelopment (testing validation and evaluation processes)[27] Importing such a prediction model in an existing or thedevelopment of a new decision support system results in asupporting tool for public authorities and citizens in choosingthe appropriate protection measures on time

21 Development of Data-Driven ANN Prediction Models Asa continuation of previous authorsrsquo research [14 15] a newprediction model based on ANN will be designed and devel-oped for the wind speed prediction as a base for the EWSNowadays the biggest problem in the development of theprediction models is their diversity of different approaches tothemodeling process their complexity procedures formodelvalidation and evaluation and implementation of differentand imprecise methodologies that can in the end leadto practical inapplicability Therefore the aforementionedmethodology [14 15] is developed on the basis of the generalmethodology guidelines suggested by Maier et al (2010) [28]and consists of the precise procedural stepsThese steps allowimplementation of the model on the new case study with adefined approach to complete modeling process

Within this paper steps of the methodology for thedevelopment of the ANN model are going to be brieflydescribed while in the dissertation done by Susanj (2017) [15]the detailed description of the methodology can be foundThedevelopedmethodology consists of the fourmain process

groups (i) monitoring (ii) modeling (iii) validation and (iv)evaluation

211 Monitoring The first process group of the aforemen-tioned methodology is the monitoring group which refers tothe procedures of collecting relevant data (both historical anddata from monitoring) and how it should be conducted It isvery important that the specific amount of the relevant andaccurate data is collected as well as the triggering factors forthe purpose of the EWS development to be recognized

212 Modeling The second group refers to the modelingprocess comprising the implementation of the ANN Inthis process identification of the model input and outputdata has to be determined Additionally data preprocessingelimination of data errors and division of data into trainingvalidation and evaluation sets have to be done When data isprepared the implementation of the ANN can be conducted

In the proposed methodology [14 15] the implemen-tation of the Multiple Layer Perceptron (MLP) architecturewith three layers (input hidden and output) is recom-mended The input layer should be formed as a matrixof a meteorological time series data multiplied by weightcoefficient 119908119896 that is obtained by learning algorithm intraining process The data should in a hydrological andmeteorological sense have an influence on the output dataTherefore it is very important that the inputmatrix is formedfrom at least ten previously measured data or one hour ofmeasurements in each line of the matrix in order to allow themodel insight into the meteorological condition in a giventime period The hidden layer of the model should consist ofthe ten neurons The number of the neurons can be changedbut the previous research has shown that it will not necessarylead to model quality improvementThe output layer consistsof the time series data that are supposed to be predicted Themodel developer chooses the timeprediction step afterwhichthe process of the ANN training validation and evaluationcan be conducted

The quality and learning strength of the ANN model isbased on the type of the activation functions and trainingalgorithm that are used [29] Activation functions are direct-ing data between the layers and the training algorithmhas thetask of optimizing the weight coefficient119908119896 in every iterationof the training process in order to provide a more accuratemodel response It is recommended to choose nonlinearactivation function and learning algorithm because of theiradaptation ability on the nonlinear problems Thereforebipolar sigmoid activation functions and the Levenberg-Marquardt learning algorithm are chosen The schematicpresentation for the ANN model development is shown inFigure 2

After the architecture of the ANN model is defined andthe training algorithm and prediction steps are chosen theprogram packages in which the model will be programmedshould be selected There are a large number of the preparedprogram packages with prefabricated ANN models but forthis purpose completion of the whole programing processis advisable Therefore usage of the MATLAB (MathWorksNatick Massachusetts USA) or any other programing soft-ware is recommended After the model is programed the

4 Complexity

1

2

10

Hidden layerInput layer Output layer

Wind speedNeuronsMeasured data

Wind speed

Levenberg-Marquardtalgorithm

Sigmoidalbipolaractivationfunction

Linearactivationfunction

Training process

matrix m times 104 matrix m times 1 matrix m times 1Meteorological data M

x1

x2

x3

x4

xm

1 = x1 times w1(k+1)




m = xm times wm(k+1)

o1(k+1) = t + Δt

o2(k+1) = t + Δt

o3(k+1) = t + Δt

o4(k+1) = t + Δt

om(k+1) = t + Δt

d1(k+1) = t + Δt

d2(k+1) = t + Δt

d3(k+1) = t + Δt

d4(k+1) = t + Δt

dm(k+1) = t + Δt

Mt12=tminus60 Mt3=tminus15Mt2=tminus10Mt1=tminus5Mt=0




Mt12=tminus60 Mt3=tminus15Mt2=tminus10Mt1=tminus5Mt=(mminus1)times5

w1(k+1) w2(k+1) w3(k+1) w4(k+1) m(k+1) w

Figure 2 Schematic presentation of the artificial neural network prediction model (119909123119898 input matrix data119872119905=(119898minus1)times5 input data setof meteorological data (wind speed wind direction wind run high wind speed high wind direction air temperature air humidity and airpressure) 119908123119898 weight coefficient V119896 sum of input matrix and weight coefficient products in 119896th iteration of the training process 119900119896neuron response for 119905 = Δ119905 in 119896th iteration of the training process 119889119896 measured data)

training process through the iterations of the calculationshould be conducted

213 Validation The third group of the model developmentprocess refers to the model validation processThis is definedas the modelrsquos quality response as the training process iscompleted The model should be validated with an earlierprepared set of the input and output data for that purpose(15 of data) in order to compare the model response withthe measured data [29 30] The model response based onthe validation data set has to be evaluated graphically andby applying numerical quality measures which are going tobe appraised according to each used numerical model qualitycriteriaTherefore it is advisable to use at least two numericalquality measures (i) Mean Squared Error (MSE) and (ii)Coefficient of Determination (1199032)

214 Evaluation The last group of the methodology stepsrefers to the model evaluation process which is defined asthe modelrsquos response quality on the data set that is not usedin the training or validation process The model should beevaluated with an earlier prepared set of the input and outputdata for that purpose (15 of data) in order to comparethe model response with the measured data [29 30] Theprocess of the model evaluation is similar to the process ofthe validation the only difference is the number of numericalquality measures that are used in that process In general itis recommended [31] to use the following measures (i) MeanSquared Error (MSE) (ii) RootMean Squared Error (RMSE)(iii)MeanAbsolute Error (MAE) (iv)Mean Squared RelativeError (MSRE) (v) Coefficient of Determination (1199032) (vi)Index of Agreement (119889) (vii) Percentage to BIAS (PBIAS)

and (viii) Root Mean Squared Error to Standard Deviation(RSR)


31 Location of the Research Area The University of Rijekacampus area (hereinafter ldquoCampusrdquo) is located in the easternpart of the city of Rijeka Primorje-Gorski Kotar CountyRepublic of Croatia and has an overall area of approxi-mately 28 ha At the location of the Campus permanentlyor occasionally stays around five thousand people (studentsand University employees) on a daily basis [33] Accordingto the specific geographical location the Campus is knownas a micro location affected by strong wind widely knownby name of the Bura It blows from the land toward the sea(Adriatic coastal area) mainly from the northeast and by itsnature is a strong wind (blowing in gusts) It usually blows forseveral days and it is caused by the spilling of cold air fromthe Pannonian area across the Dinarid mountains towardcoastline

Since 2014 theCampus area has been affected by thewindBura several times causing damage to the Campus propertyand environment as is shown in Figure 3

Because of flying objects that can cause damage to theproperty and the environment and hurt people and thepowerful gusts of the wind Bura making it impossible forpeople to move outdoors the ldquoprocedure plan in a case ofthe emergency caused by stormy wind or hurricane snowand occurrence of the ice on the University of Rijeka campusrdquo(hereinafter ldquoprocedure planrdquo) was adopted in 2015

32 Present State This procedure planwas enacted accordingto the natural disaster protection law (NN 731997) and

Complexity 5

(a) (b)

(c) (d)

Figure 3 Damage on the Campus property and environment caused by the wind Bura (a) and (b) damage inMarch 2015 (c) and (d) damagein January 2017

the Campus Technical Services (CTS) is in charge of theprocedure conduction CTS has an obligation to monitor (i)wind speed andwind direction on the Campus area (installedmeasuring equipment) and (ii) the prediction of the Croatianofficial alerting system through the METEOALARM (Alert-ing Europe for extreme weather) obtained by the Network ofEuropean Meteorological Services (EUMETNET) Accord-ing to the prediction alerting system and monitoring CTSdeclares two levels of alert (Yellow and Red)

In the case of the stormy wind that is classified as windwith ten minutesrsquo mean speed above 8 on a Beaufort scale(V gt 189ms) the CTS will pronounce the first level of alertas ldquoYellow alert stage of preparation for the emergencyrdquo TheCTS will pronounce the second level of alert defined as ldquoRedalert extreme danger state of the natural disasterrdquo in the caseof a hurricane wind when the ten minutesrsquo mean wind speedreaches 10 on a Beaufort scale (V gt 25ms)

In the case of a Yellow alert it is recommended to closethe windows on the buildings and not be in the vicinityof possible flying objects and open windows In the case ofa Red alert the area of the Campus is closed (postponingteaching and research activities) and people are not allowedto move outside of the buildings The announcements aredisseminated by e-mail to all University employees throughthe local radio stations and on the official Internet web pagesof every University constituent The existing system alert isshown in Figure 4

The implemented procedure plan is very simple and ithas several deficiencies As the procedure is not automated

Campus technicalservices

Meteoalarm

Alert stages

Y R

Meteorologicalstation(on Campus)

Dissemination ofalerts

Figure 4 Scheme of the ldquoprocedure plan in a case of the emergencycaused by stormy wind or hurricane snow and occurrence of the iceon the University of Rijeka campusrdquo

the dissemination of alerts is based on METEOALARM andthe alertness of CTS employees to disseminate informationBigger problems occur in METEOALARMrsquos macro level ofprediction as the possible occurrence of strong wind is notsufficiently accurate on the micro level due to the smallnumber of meteorological stations in the region and versatilerelief The University of Rijeka campus area is placed on thespecific micro location on which the speed of the wind gusts

6 Complexity

Prediction model(ANN)


Meteoalarm Meteorologicalstation(on Campus)

Decision support


Y R

Alert stages

Y R

Δt = 1

Δt = 1

Δt = 24

Δt = 1

Δt = 1 Δt = 24

Δt = 24

Δt = 24

Figure 5 Scheme of the proposed EWS

is usually greater than that in the rest of the wider Rijeka cityarea The continuous monitoring and collection of the windspeed and direction data are implemented on the Campusarea but with those measurements it is only possible to seepast and real-time meteorological variables such as windspeed and wind direction data In this case the real-timemeasurements are useless to the CTS in order to announcealerts on time and they can serve only to observe the currentstate The aforementioned location weather conditions implythat it is possible to have a stronger wind on Campus thanthat predicted by METEOALARM which can potentiallyhave a hazardous impact on both property and people Thedissemination of the announcements is also aweak part of theexisting system because of the large number of the studentsthat are not directly alerted

Therefore it is necessary to improve the existing alertingsystem by the implementation of the EWS that is going tobe based on both the METEOALARM and the wind speedprediction model and also to improve the procedure plan bythe development of the dissemination procedures to obtainbetter response capability As the aim and objective is todesign and develop an ANN model in order to achieve asuccessful prediction of wind speed on micro location thefocus of the proposed EWS (Figure 5) is on overlappingmicrolevel prediction with macro level forecast

As stated a proposed procedure plan is based onMETEOALARMrsquos alert stages and real-time processed datafromon-Campusmeteorological station Collected data fromthe meteorological station serve as input for ANN model

which gives results in the form of predicted wind speed inseveral time prediction steps This provides an opportunityto the decision-maker to overlap predictions of ANN modelof micro level with macro levels stages of alert in order todisseminate alerts toward stakeholders for specific locationsThereforemicro levelmodels can give an in-depth predictionon locationThis is of great importance when there is a Yellowalert (on the macro level) and the ANN model predicts thatconditions on the micro level are Red

33 Data Collection As mentioned before in the method-ology for the development of the model the first group ofsteps refers to the establishment of themonitoringThereforecontinuous data monitoring of the meteorological variableshas been established since the beginning of 2015 Data usedfor modeling purposes dated from January 1 2015 to June1 2017 Meteorological variables used for prediction are (i)wind speed (ii) wind direction (iii) wind run (iv) high windspeed (v) high wind direction (vi) air temperature (vii)air humidity and (viii) air pressure They were collected byVantage Pro 2meteorological station (manufactured byDavisInstruments Corporation) that is installed on the roof of theFaculty of Civil Engineering The measurement equipmentis obtained by the RISK project (Research Infrastructurefor Campus-Based Laboratories at the University of RijekaRC2206-0001)The frequency of datameasurement intervalis five minutes

34 Development of ANNWind Prediction Model The afore-mentioned collected data is used for the development of thewind speed prediction model for the University of Rijekacampus area Input (8 variables) and output (1 variable)of the model are selected and then preprocessed Data isthen divided into the sets for the training (70 of totaldata) validation (15 of total data) and evaluation (15 oftotal data) of the model Statistics of data used for trainingvalidation and evaluation processes are shown in Table 1

After data is prepared implementation of theANNmodelis conducted by MATLAB software (MathWorks NatickMassachusetts USA) and the schematic representation of themodel is shown in Figure 6

Model training is conducted for the time prediction steps(i) Δ119905 = 1 h (ii) Δ119905 = 3 h (iii) Δ119905 = 8 h (iv) Δ119905 = 12 hand (v) Δ119905 = 24 h After it is trained the model is visuallyand numerically validated and evaluated A comparison ofthe measured and predicted wind speed in the process ofevaluation according to time prediction step is presented inFigure 7 On the basis of the visual analysis it can be observedthat the accuracy of the model decreases according to theextension of the prediction time step as expected

Numerical validation and evaluation of the modelare conducted according to the proposed aforementionedmethodology Results of the validation and evaluation pro-cesses and assessment of the model quality are shown inTable 2 Since some of the numerical model quality measurescriteria are not precisely determined (measures MSE RMSEMAE MSRE and RSR) it is recommended that the resultsshould be close to zero For others used measures criteria areprecise and models are assessed according to them

Complexity 7

Table 1 Statistics of data used for training validation and evaluation of the ANN wind prediction model

StatisticslowastInput layer Output layer

Windspeed

Winddirection Wind run High wind

speedHigh winddirection

Airpressure

Airhumidity

Airtemperature Wind speed

[ms] [ ] [km] [ms] [ ] [hPa] [] [∘C] [ms]Model training data (70 of data)

119899 156552 156552 156552 156552 156552 156552 156552 156552 156552Max 259 NNW 1127 46 NNE 10369 98 414 259Min 0 0 0 9725 11 minus5 0120583 268 083 498 101574 5995 1536 268120590 253 082 442 2116 82 253

Model validation data (15 of data)119899 33547 33547 33547 33547 33547 33547 33547 33547 33547Max 219 NNE 657 38 NNE 10394 98 271 219Min 0 0 0 9955 7 minus81 0120583 265 08 507 102149 6514 818 265120590 235 07 45 821 2168 563 235

Model evaluation data (15 of data)119899 33547 33547 33547 33547 33547 33547 33547 33547 33547Max 161 NNE 483 246 N 10347 97 371 161Min 0 0 0 9979 14 0 0120583 214 065 417 10167 5786 1602 214120590 158 047 296 606 2036 771 158lowast119899 = number of observations Max = maximum Min = minimum 120583 = sample mean 120590 = standard deviation

Table 2 Performance statistics of the ANN model during validation and evaluation processes

Δ119905[h]

Validation EvaluationMSE 1199032 MSE RMSE MAE MSRE 1199032 119889 PBIAS RSR[(ms)2] [-] [(ms)2] [ms] [ms] [-] [-] [-] [] [-]

1 0075 0993 0058 0158 0113 2334 0987 0998 0503 01053 2033 0951 1505 0736 0581 124898 0894 0963 6817 04818 2999 0868 1970 1094 0897 4365318 0727 085 14969 072912 3872 0833 2306 1216 1026 316953 0669 0727 17302 079424 4985 0572 2741 1382 1088 1116616 0467 0353 9539 0923Quality Criteria Very good in bold good in italic and poor in bold italic Model quality criteria according to [31 32]

Numerical validation and evaluation measures have con-firmed that themodels with time stepsΔ119905 = 1 hΔ119905 = 3 h andΔ119905 = 8 h have noticeably better prediction possibilities sincethey are assessed as ldquovery goodrdquo and ldquogoodrdquo than modelswith time steps Δ119905 = 12 h and Δ119905 = 24 h Additionallyother measures are very small near zero while for the timesteps Δ119905 = 12 h and Δ119905 = 24 h they are significantly biggerAlthough models with time steps Δ119905 = 12 h and Δ119905 = 24 hshow poor quality results according to visual analysis theyare still showing some prediction indication of an increase ordecrease of wind speed and therefore these predictions canbe used but carefully

According to the results of performance statistics of thedeveloped ANN model the decision-maker on micro levelcanmake decisions on timewith up to prediction step ofΔ119905 =8 h By overlapping the macro level forecast with this microlevel prediction they can control the accurate disseminationof alerts in a specific area

4 Conclusions

This paper presents an application of artificial neural net-works in the predicting process of wind speed and itsimplementation in earlywarning systemas a decision support

8 Complexity

MeasurementOutput layerHidden layer

10neurons

Data matrix

Wind speedWind directionWind runHigh wind speedHigh wind directionAir temperatureAir humidityAir pressure

Wind speedprediction Wind speed


Input layer

Data matrix

Data matrix

Model response

Model response

ANN model

(33547 times 104 data)


(156552 times 104 data)

Δt = 1 hour

Δt = 3 hours

Δt = 8 hours

Δt = 12 hours

Δt = 24 hours

Trainingprocess

Comparison modelresponse andmeasured data



Validationprocess

Evaluationprocess

Visual andnumericalvalidationmeasures


Figure 6 Schematic representation of the ANN wind speed prediction model based on [14 15]

1834 1836 1838 184 1842 1844 1846 1848 185Time (minutes)

2468

1012141618

Win

d sp

eed

(ms

)

Measurement

times105

Δt = 1 hourΔt = 3 hours

Δt = 8 hoursΔt = 12 hoursΔt = 24 hours

Figure 7 Comparison of the measured wind speed and modelresponse (fiveminutesrsquo time step) in process of themodel evaluationaccording to time prediction steps (Δ119905 = 1 h Δ119905 = 3 h Δ119905 = 8 hΔ119905 = 12 h and Δ119905 = 24 h)

tool The main objectives of this research were to develop themodel to achieve a successful prediction of wind speed onmicro location based on data frommeteorological station andto implement it in model base to serve as decision supporttool in the proposed early warning system

Data gathered by a local meteorological station duringa 30-month period (8 variables) was used in the predictionprocess of wind speed The model is developed for 5-timeprediction steps (i) Δ119905 = 1 h (ii) Δ119905 = 3 h (iii) Δ119905 = 8 h(iv) Δ119905 = 12 h and (v) Δ119905 = 24 h The evaluation of themodel shows very good prediction possibilities for time stepsΔ119905 = 1 h Δ119905 = 3 h and Δ119905 = 8 h and therefore enables theearly warning system implementation

The performed research indicates that it is possible anddesirable to apply artificial neural network for the predictionprocess on the micro locations because that type of model isvaluable and accurate tool to be implemented in model baseto support future decisions in early warning systems

Complexity 9

The complete evaluation and functionality of the pro-posed early warning system and artificial neural networkmodel can be done after its implementation on theUniversityof Rijeka campus (Croatia) is conducted



Acknowledgments

The research for this paper was conducted within the projectldquoResearch Infrastructure for Campus-Based Laboratories atthe University of Rijekardquo which is cofunded by the EuropeanUnion under the European Regional Development Fund(RC2206-0001) as well as a part of the scientific projectldquoWater ResourcesHydrology and Floods andMudFlowRisksIdentification in the Karstic Areardquo financed by the Universityof Rijeka (13051103)

References

[1] A De Bono B Chatenoux C Herold and P Peduzzi GlobalAssessment Report onDisaster Risk Reduction 2013 From SharedRisk to Shared Decision Support for Risk Management Designof EWS 13 Value-The Business Case for Disaster Risk ReductionUNISDR Geneva Switzerland 2013

[2] Harvard Humanitarian Initiative et al Disaster Relief 20 ldquoThefuture of information sharing in humanitarian emergenciesrdquoHHI United Nations Foundation OCHATheVodafone Foun-dation Washington DC USA 2010

[3] J C Gaillard and J Mercer ldquoFrom knowledge to actionBridging gaps in disaster risk reductionrdquo Progress in HumanGeography vol 37 no 1 pp 93ndash114 2013

[4] J Weichselgartner and R Kasperson ldquoBarriers in the science-policy-practice interface Toward a knowledge-action-system inglobal environmental change researchrdquo Global EnvironmentalChange vol 20 no 2 pp 266ndash277 2010

[5] E L Quarantelli ldquoDisaster crisis management a summary ofresearch findingsrdquo Journal of Management Studies vol 25 no4 pp 373ndash385 1988

[6] UNISDR platform for the promotion of early warning (PPEW)and un Secretairat of the International strategy for disasterreduction (UNISDR) ldquoDeveloping early warning system Achecklistrdquo in Proceedings of the EWC III Third InternationalConference on Early Warning From Concept to Action BonnGermany March 2006

[7] L Lin J T Eriksson H Vihriala and L Soderlund ldquoPredictingwind behavior with neural networksrdquo in Proceedings of the1996 EuropeanWind Energy Conference pp 655ndash658 GoteborgSweden May 1996

[8] M A Mohandes S Rehman and T O Halawani ldquoA neuralnetworks approach for wind speed predictionrdquo Journal ofRenewable Energy vol 13 no 3 pp 345ndash354 1998

[9] M C Alexiadis P S Dokopoulos H S Sahsamanoglou andI M Manousaridis ldquoShort-term forecasting of wind speed andrelated electrical powerrdquo Solar Energy vol 63 no 1 pp 61ndash681998

[10] P M Fonte G X Silva and J C Quadrado ldquoWind speed pre-diction using artificial neural networksrdquo WSEAS Transactionson Systems vol 4 no 4 pp 379ndash383 2005

[11] MMonfared H Rastegar and HM Kojabadi ldquoA new strategyfor wind speed forecasting using artificial intelligent methodsrdquoJournal of Renewable Energy vol 34 no 3 pp 845ndash848 2009

[12] S A Pourmousavi Kani and M M Ardehali ldquoVery short-termwind speed prediction a new artificial neural network-Markovchain modelrdquo Energy Conversion and Management vol 52 no1 pp 738ndash745 2011

[13] Q Zhou T Xiong M Wang C Xiang and Q Xu ldquoDiagnosisand early warning of wind turbine faults based on clusteranalysis theory and modified ANFISrdquo Energies vol 10 no 7article 898 2017

[14] I Susanj N Ozanic and IMarovic ldquoMethodology for develop-ing hydrological models based on an artificial neural networkto establish an early warning system in small catchmentsrdquoAdvances in Meteorology vol 2016 Article ID 9125219 14 pages2016

[15] I Susanj Development of the hydrological rainfall-runoff modelbased on artificial neural network in small catchments [PhDthesis] University of Rijeka Faculty of Civil EngineeringRijeka Croatia 2017

[16] AAsemi A Safari andAAsemiZavareh ldquoThe role ofmanage-ment information system (MIS) and decision support system(DSS) for managerrsquos decision making processrdquo InternationalJournal of Business and Management vol 6 no 7 pp 164ndash1732011

[17] I Marovic Decision support system in real estate value man-agement [PhD thesis] University of Zagreb Faculty of CivilEngineering Zagreb Croatia 2013

[18] N Jajac Design of decision support systems in the managementof infrastructure systems of the urban environment [MS thesis]University of Split Faculty of Economics Split Croatia 2007

[19] N Jajac S Knezic I Marovic S Knezic and I MarovicldquoDecision support system to urban infrastructure maintenancemanagementrdquo Organization Technology amp Managament inConstruction vol 1 no 2 pp 72ndash79 2009


[21] M H GlantzUsable Science 8 Early Warning Systems Dorsquos andDonrsquots Report of Workshop Shanghai China October 2003

[22] Millennium Ecosystem Assessment Ecosystems and HumanWell-Being A Framework for Assessment Island PressWashing-ton DC USA 2003

[23] P A Schrodt and D J Gerner ldquoThe impact of early warningon institutional responses to complex humanitarian crisesrdquo inProceedings of the Third Pan-European International RelationsConference and Joint Meeting with the International StudiesAssociation Vienna Austria September 1998

[24] P Hall ldquoEarly warning systems Reframing the discussionrdquoAustralian Journal of Emergency Management vol 22 no 22007

[25] UNISDR International Strategy for Disaster ReductionldquoHyogo framework for action 2005ndash2015 building the resilienceof nations and communities to disastersrdquo in Proceedings ofthe World Conference on Disaster Reduction Hyogo JapanJanuary 2005

[26] T Comes B Mayag and E Negre ldquoDecision support fordisaster riskmanagement integrating vulnerabilities into early-warning systemsrdquo in Proceedings of the International Conferenceon Information Systems for Crisis Response and Managementin Mediterranean Countries pp 178ndash191 Springer InternationalPublishing Cham Germany October 2014

10 Complexity

[27] V V Krzhizhanovskaya G S Shirshov and N B MelnikovaldquoFlood early warning system design implementation andcomputational modulesrdquo Procedia Computer Science vol 4 pp106ndash115 2011

[28] H R Maier A Jain G C Dandy and K P Sudheer ldquoMethodsused for the development of neural networks for the predictionof water resource variables in river systems current status andfuture directionsrdquo Environmental Modeling and Software vol25 no 8 pp 891ndash909 2010

[29] H B Demuth and M H Beale Neural Network ToolboxTMUserrsquos Guide The MathWorks INC 2004

[30] M T Hagan H B Demuth M H Beale O De Jes and O DeJesus Neural Network Design (20) PWS Publishing CompanyBoston Mass USA 1996

[31] R Abrahart P E Kneale and L M See Neural Networks forHydrological Modelling Taylor amp Francis Group London UK2004

[32] P Matic Short-term forecasting of hydrological inflow by use ofthe artificial neural networks [PhD thesis] University of SplitFaculty of Electrical EngineeringMechanical Engineering AndNaval Architecture Split Croatia 2014

[33] S Grbac L Malnar A Vorkapic D Car-Pusic and I MarovicldquoldquoPreliminary analysis of the spatial attributes affecting stu-dentsrsquo quality of life at the University of Rijekardquo in Proceedingsof the People Buildings and Environment 2016 an InternationalScientific Conference vol 4 pp 109ndash117 Luhacovice CzechRepublic 2016

Research ArticleEstimation of Costs and Durations of Construction ofUrban Roads Using ANN and SVM

Igor Peško Vladimir MuIenski Miloš Šešlija Nebojša RadoviT Aleksandra VujkovDragana BibiT and Milena Krklješ

University of Novi Sad Faculty of Technical Sciences Trg Dositeja Obradovica 6 Novi Sad Serbia

Correspondence should be addressed to Igor Pesko igorbpunsacrs

Received 27 June 2017 Accepted 12 September 2017 Published 7 December 2017

Academic Editor Meri Cvetkovska

Copyright copy 2017 Igor Pesko et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Offer preparation has always been a specific part of a building process which has significant impact on company business Due tothe fact that income greatly depends on offerrsquos precision and the balance between planned costs both direct and overheads andwished profit it is necessary to prepare a precise offer within required time and available resources which are always insufficientThe paper presents a research of precision that can be achieved while using artificial intelligence for estimation of cost and durationin construction projects Both artificial neural networks (ANNs) and support vector machines (SVM) are analysed and comparedThe best SVMhas shown higher precision when estimating costs withmean absolute percentage error (MAPE) of 706 comparedto the most precise ANNs which has achieved precision of 2538 Estimation of works duration has proved to be more difficultThe best MAPEs were 2277 and 2626 for SVM and ANN respectively

1 Introduction

Civil engineering presents a specific branch of industry fromall aspectsThemain reason for this lies in specific features ofconstruction objects as well as the conditions for their realisa-tion Another specific aspect of realisation of constructionprojects is that the realisation process involves a large numberof participants with different rolesThe key role in the realisa-tion of construction processes is certainly played by aninvestor who is at the same time the initiator of realisation of aconstruction project whose main goal is to choose a reliablecontractor who can guarantee the fulfilment of set require-ments (costs time and quality)

When choosing a constructor a dominant parameteris the offered cost of realisation which implies that it isnecessary to carry out an adequate estimation of constructioncosts The question that arises is which costs that is whichprice should be the subject of estimation Gunner andSkitmore [1] and Morrison [2] based their research whichrelates to the estimation of realisation costs on the lowestprice offered However according to Skitmore and Lo [3]very often the lowest price does not reflect actual realisationcosts since contractors offer services at unrealistically low

prices Azman et al [4] quote the recommendation of Loweand Skitmore to accept the second lowest offer as the verylowest one does not guarantee the actual value They alsomention the suggestion made by McCaffer to use the meanvalue of submitted offers for estimation rather than thelowest one explaining that it is the closest to the actual price

Some authors however Aibinu and Pasco [5] use thevalues given in the accepted offer for the needs of estimatingthe accuracy of predicted values since it is the value acceptedby the investor Abou Rizk et al [6] and Shane et al [7]recommend the use of actually paid realisation value Never-theless Skitmore [8] explains that there are certain problemswhen using actually paid values for realisation with the mainone residing in the fact that the data is not easily available andeven if they are they are often not properly recorded that isare not real Moreover there is a problem of time betweenthe forming of cost estimation and realisation of real costsin which significant changes in the project may occur duringthe realisation process which can influence the amount ofcontracted works for which the offer was formed

According to everythingmentioned above the estimationof costs within this research was carried out based on therealisation value offered by the contractor


2 Complexity

Further on there are two levels of estimation of potentialworks from the perspective of a contractor which precede therealisation of contracting and those are conceptual (rough)and preliminary (detailed) estimation [9] Conceptual esti-mation of costs usually results in the total number withoutthe detailed analysis of the structure of costs This impliesthat conceptual estimation should use simpler and specialisedestimation models Its contribution is the assessment ofjustifiability of following work on the project in question ormore precisely further work on preliminary estimationwhichis carried out prior to signing of a contract The basis forboth estimations is the data about the object provided by theinvestor that is tender documentation The question ariseson what accuracy of estimation is acceptable The requiredacceptable accuracy of estimation of construction costs fromthe perspective of a contractor in the initial phase of tenderprocedure (conceptual estimation) according to Ashworth[10] amounts toplusmn15This is the accuracywhichwas adoptedin this research as the basic goal of precision of formedmodels for the estimation of construction costs

It was noted that the dominant parameter for choosingthe most favourable contractor is the offered price Howeverthe proposed time for realisation of works in question shouldnot be neglected either The main problem of estimation ofduration of works in the conceptual phase is the fact that thepotential contractor does not possess realisation plans whichimplies the application of estimation methods which providesatisfactory accuracy based on data available at a time It is acommon case for an investor to limit the maximum durationof construction in the tender conditions

Having in mind that it is necessary to carry out simulta-neous estimation of costs and duration of the constructionprocess the fact that the research confirmed that the costs ofsignificant position of works are at the same time of temporalsignificance is highly important [11 12] The previously men-tioned statement is significant for the processes of planningand control of time during the realisation of a constructionproject and confirms the justifiability of simultaneous estima-tion of costs and duration of contracted works

It was mentioned that conceptual estimation requiresapplication of simpler and specialisedmodels with an accept-able accuracy of estimation One of the possible approachesis the application of artificial intelligence (artificial neuralnetworks (ANNs) support vector machine (SVM) etc) Thebasic precondition for the application of this approach is theforming of an adequate base of historical data on similarconstruction projects previously realised

2 ANN and SVM Applications inConstruction Industry

Thefirst scientific article related to the application of ANNs inconstruction industry was published by Adeli in the Micro-computers in Civil Engineering magazine [13] The applica-tion becomes more frequent along with software develop-ment The sphere of application of ANNs and the SVM inconstruction industry is very wide and present in basically allphases of realisation of a project from its initialising throughdesigning and construction to maintenance renovation

demolishing and recycling of objects Some of the examplesof their application are estimation of necessary resources forrealisation of projects [14] subcontractor rating [15] estima-tion of load lifting time by using tower cranes [16] noting thequality of a construction object [17] application in carryingout of economic analysis [18] application in the case of com-pensation claims [19 20] construction contractor defaultprediction [21] prequalification of constructors [22] cashflow prediction [23] project control forecasting [24] esti-mation of recycling capacity [25] predicting project perfor-mance [26] and many others

Cost estimation by using ANNs and SVM is often repre-sented in the literature for example estimation of construc-tion costs of residential andor residential-commercial facili-ties [27 28] cost estimation of reconstruction of bridges [29]and cost estimation of building water supply and seweragenetworks [30] In his master thesis Siqueira [31] addressesthe application of ANN for the conceptual estimation of con-struction costs of steel constructions An et al [32] showedthe application of SVM in the assessment process of construc-tion costs of residential-commercial facilities In addition tothementioned authors the application of ANNs and SVM forthe estimation of construction costs was dealt with by Adeliand Wu [33] Wilmot and Mei [34] Vahdani et al [35] andmany others

Kong et al [36] carried out the prediction of cost per m2for residential-commercial facilities by using SVM Kong etal [37] carried out the comparison of results of price predic-tion for the same data base by using a SVM model and RS-SVMmodel (RS rough set)

Kim et al [38] carried out a comparative analysis of resultsobtained through ANN SVM and regression analysis forcost estimation of school facilities This presents a rathercommon case in available and analysed literature Not onlythe comparative analysis of ANN an SVMmodels with othermodels for solving of the same types of problems but alsoits combining with other forms of artificial intelligence suchas fuzzy logic (FL) and genetic algorithms (GA) is carriedout where the so-called hybrid models are created Kim et al[39] comparedmodels for cost estimation of the constructionof residential facilities based on multiple regression analysis(MRA) artificial neural networks (ANNs) and case-basedreasoning (CBR) Sonmez [40] also drew a parallel betweenan ANN model for conceptual estimation of costs with aregression model Kim et al [41] and Feng et al [42] usedgenetic algorithms (GA) in their research when defining anANNmodel for cost estimation of construction of buildingsas a tool for optimisation of the ANNmodel itself In additionto combining withGA there is also the possibility of combin-ing an ANN with fuzzy logic (FL) where hybrid models arecreated as wellThe application of ANN-FL hybridmodels forthe estimation of construction costs was used in the researchof Cheng and Huang [43] and Cheng et al [44] Chengand Wu [45] outlined the comparative analysis of modelsfor conceptual estimation of construction costs formed byusing of ANN SVM and EFNIM (evolutionary fuzzy neuralinference model) Deng and Yeh [46] showed the use of LS-SVM (LS least squares) model in the cost prediction processCheng et al [47] connected two approaches of artificial

Complexity 3

intelligence (fast genetic algorithm (fmGa) and SVM) for thepurpose of estimation of realisation of construction projects(eg prediction of completeness of 119894th period in the reali-sation of the construction project)This model is called ESIM(evolutionary support vector machine inference model)

Since the topic of the research is the estimation of costsand duration of construction of urban roads the followingtext will contain research carried out on similar types ofconstructions Wang et al [48] showed in their work theestimation of costs of highway construction by using ANNsIncluding the set of 16 realised projects they carried outthe training (first 14 projects) and validation (remaining 2projects) of the ANNmodel Al-Tabtabai et al [49] surveyedfive expert projectmanagers defining the factors which influ-ence the changes in the total costs of highway constructionsuch as location maintaining of the existing infrastructuretype of soil consultantrsquos capability of estimation constructionof access roads the distance of material and equipmenttransportation financial factors type of urban road and theneed for obtaining of a job

Hegazy andAyed [50] provided an overview of forming ofa model by using ANNs for the parameter estimation of costsof highway construction They use the data obtained from18 offers by anonymous bidders for the realisation of workson highway construction in Canada 14 offers for trainingand 4 offers for testing Sodikov [51] gave an outline ofthe estimation of costs of highway construction by applyingANNs The analysis and formation of model were performedbased on two sets of data from different locationsThe first setof data was formed based on the projects realised in Poland(the total of 135 projects with 38 realised projects on highwayconstruction used for the analysis) and the second based onprojects realised inThailand (the total of 123 projects with 42realised projects on asphalt layer ldquocoatingrdquo) It is important tonote that Hegazy and Ayed [50] as well as Sodikov [51] usedthe duration of works realisation as the input parameterThisduration is also unknown in the process of defining of theconceptual estimation as is the cost of realisation of worksEl-Sawahli [52] performed the estimation of costs of roadconstruction by using a SVMmodel formed according to thebase of 70 realised projects in total

Estimation of duration of the construction of buildingsby using ANNs and SVM is not present in the literature as itis the case with estimation of costs Attal [53] formed inde-pendent and separate ANN models for estimation of costsand duration of highway construction Bhokha andOgunlana[54] defined an ANN model for prediction of duration ofmultistorey buildings construction in a preproject phaseHola and Schabowicz [55] carried out the estimation of dura-tion and costs of realisation of earthworks by using an ANNmodelWang et al [56] performed the predicting of construc-tion cost and schedule success by applying ANN and SVM

3 Material and Methods

Within the research carried out gathering of data and dataanalysis were performed followed by the preparation of datafor the needs of model formation as well as the final formingof models and their comparative analysis

Gathering of data and forming of the data base on realisedconstruction projects of reconstruction andor constructionof urban roads were carried out on the territory of the city ofNovi Sad the Republic of Serbia All the projects were fundedby the same investor in the period between January 2005and December 2012 All the projects solely relate to the real-isation of construction works based on completed projectsand technical documentation Having in mind everythingmentioned above uniform tender documentation that is thebill of works whichmakes its integral part served as themainsource of information

The sample comprises 198 contracted and realised con-struction projects However not all of the projects wereincluded in the analysis carried out later on owing it tothe fact that certain projects 32 of them in total includerealisation of works on relocation of installations (seweragewater gas etc) green spaces landscaping street lighting andconstruction of supporting facilities on the roads (smallerbridges culverts etc) which are excluded from furtheranalysis

The total number of realised projects included in thefurther analysis amounts to 166 projects of basic constructionworks andor reconstruction of urban roads As it has beenalready noted all projects were realised for the same investorConsequently the tender documentation is uniform whichis also the case with the distribution of works within thebill of works dividing them into preparation works earth-works works on construction of pavement and landscapingdrainage works works on construction of traffic signals andother works

The analysis of the share of costs of the mentionedwork groups in the total offered price of realisation wascarried outThis confirmed that the works on roadway struc-ture and landscaping have the biggest impact on the totalprice ranging within the interval from 2224 to 100(only two cases where only these two types of works wereplanned) Earlier research included similar analysis but onlyfor projects realised by the same contractor [57 58]

Table 1 shows mean values of the percentage of worksgroups in the total offered value for realisation of works Theinterval between 65 and 95 of the amount of works onthe roadway construction and landscaping within the totaloffered price includes 106 projects that is 6386 of the totalnumber of analysed projects The total number of potentialwork positions which may occur during the realisation of allbasic works is 91 (according to the general section of tenderdocumentation which is the same for all analysed projects)whereas the total number of potential positions from the billof works related to roadway construction and landscapingworks amounts to 18 Percentage share of activities related toroadway construction and landscaping in relation to the totalnumber of potential positions amounts to 1978 which isapproximately the same as the percentage of cost-significantactivities according to ldquoParetordquo distribution which amountsto 20

Considering the fact that for 6386 of analysed projectsthe share of the total price ranges within the interval betweenplusmn15 and 80 of the total offered value it can be statedthat the works on roadway construction and landscaping are

4 Complexity

Table 1 Mean values of the percentage of works groups in the total offered value

Number Group of works Mean value of the percentage of works groups in the total offered price [](1) Roadway construction and landscaping works 68(2) Earthworks 14(3) Preparation works 12(4) Other works 3(5) Drainage works 2(6) Works on traffic signals installation 1

Table 2 Number of projects according to the offered number of days for realisation

Number Offered numberof days

Number ofprojects

Percentage sharein the data base

[](1) Up to 20 50 3012(2) 21 to 30 31 1867(3) 31 to 40 24 1446(4) 41 to 50 26 1566(5) 51 to 60 14 843(6) 61 to 70 5 301(7) 71 to 80 7 422(8) 81 to 90 4 241(9) Above 90 5 301

cost-significant according to ldquoParetordquo distribution that isdistribution of 2080 This is an additional reason why theseworks play themajor rolewhen defining an estimationmodelRealisation of works on roadway construction and landscap-ing relates to usage of basic materials such as crushed stone(different fractions) curbs asphalt base layer asphalt surfacelayer and concrete prefabricated elements for paving

Duration of realisation is offered in the form of the totalnumber of days and there is noway inwhich it can be claimedwith certainty how much time is needed for the realisationof each group of works individually For this reason classifi-cation of projects was carried out based solely on the totalamount of time offered for the realisation of all plannedworks Table 2 shows the number of projects according tothe offered number of days for realisation According tosome research cost-significant work positions are duration-significant as well [11 12] Hence it is of equal importanceto put particular emphasis on works related to roadway con-struction and landscaping both from the perspective of costestimation and from the duration of the realisation process

Estimation of costs as well as all the other estimationssuch as duration of realisation involves the engagement ofresources According to the literature estimation of costsranges between 025 and 1 of the total investment value[59] For this reason Remer and Buchanan [60] developed amodel for estimating of costs for the work (cost) estimationprocessThe reason for such research lies in the fact that low-cost estimates can lead to unplanned costs in the realisationprocess andor in lower functional features of an object thanthose required by the investor

The primary purpose of forming of an estimation modelis to perform the most accurate estimation possible inthe shortest interval of time possible with the minimumengagement of resources all of it based on the data availableat the time of estimation Since the contracts in question arethe so-called ldquobuildrdquo contracts an integral part of tender doc-umentation is the bill of works or more precisely the amountof works planned in the project-technical documentationEstimations of costs based on amounts and unit prices arethe most accurate ones but require a great amount of timeand need to be applied in preliminary (detailed) estimationwhich precedes directly the signing of a contract Howeverconceptual (rough) estimation which results in the totalcosts of realisation requires simpler and faster methods ofestimationWith the aim of defining amethod featuring suchperformances a formerly presented analysis of significanceand impact of groups of works on the total price both on totalcosts and on realisation time was carried out

The works on roadway construction and landscapingas the most important group of works were considered inmore detail in comparison with other groups of works thatis they were assigned greater significance Having in mindthe characteristics of works performed within this group alarge amount of material necessary for their realisation beingone of them the input parameters for creating of this modelare the amounts of material necessary for their realisation(Table 4)

Despite the fact that the mean percentage share ofroadway construction and landscaping works amounts to68 the share of the remaining groups of works in the total

Complexity 5

Table 3 Number of projects depending on the offered revalorised price for realisation of works

Number Offered price for realisation [RSD]lowast Number of projects Percentage share in data base [](1) Up to 5000000 32 192(2) From 5000000 to 10000000 28 1687(3) From 10000000 to 20000000 24 1446(4) From 20000000 to 30000000 19 1145(5) From 30000000 to 40000000 13 783(6) From 40000000 to 60000000 8 482(7) From 60000000 to 100000000 21 1265(8) From 100000000 to 200000000 13 783(9) Over 200000000 8 482lowastRSD Republic Serbian Dinar (1 EUR = 122 RSD)

Table 4 Inputs into models

Number Description of input data Data type Unit ofmeasurement Min Max Mean

valueInput 1 Amount of crushed stone Numerical m3 000 1607000 169462Input 2 Amount of curbs Numerical m1 000 1430000 197507Input 3 Amount of asphalt base layer Numerical t 000 3156900 111961Input 4 Amount of asphalt surface layer Numerical t 000 1104600 50585Input 5 Amount concrete prefabricated elements Numerical m2 000 2000000 282485

Percentage share of wok positionsInput 6 Preparation works Numerical 000 10000 4318Input 7 Earthworks Numerical 000 10000 4892Input 8 Drainage works Numerical 000 9333 1663Input 9 Traffic signalisation works Numerical 000 10000 2866Input 10 Other works Numerical 000 10000 859Input 11 Works realisation zone Discrete - 1 2 -

Input 12 Project category (values of up to and over40000000) Discrete - 1 2 -

offered value should not be neglectedThe biggest share in thetotal offered value for realisation of the remaining groups ofworks belongs to earthworks and preparation works whereastraffic signals other works and drainage contribute with aconsiderably lower percentage (Table 2)

Inclusion of the mentioned works in further analysis wascarried out based on the number of planned positions ofworks for each individual bid project in relation to a possiblenumber of positions of works (according to a universal listissued by the investor) in groups of works directly throughpercentage share (Table 4)

Moreover it was noticed that it is possible to classifyrealisedworks based on the location theywere supposed to berealised on For that purpose realisation ofworkswas dividedinto two zones zone 1 realisation of works in the city centreand zone 2 realisation of works in the suburbs (Table 4)

Since the research carried out relates to the financialaspect of realisation of construction projects it is necessary toperform revalorisation in order for the data to be comparableand applicable to forming of an estimation model by usingartificial intelligence By using the revalorisation processdefining of difference is achieved that is the increase or

decrease of offered values for the realisation of works inrelation to the moment in which the estimation of realisationcosts for future projects will be made by applying the modelIn other words by applying the revalorisation process thechanges in the prices defined on the base date in relationto the current date will be defined Base date is the dateon which the contracted price was formed (ie giving anoffer) whereas the current date presents the date on whichthe revalorisation is carried out

Revalorisation ismade based on the index of general retailprices growth in the Republic of Serbia where the increasein the period between February 2005 and July 2012 amountedto 9547 which is close to the mean value of increase of8991 obtained on the basis of the values of unit pricesfrom two final offers After the completed revalorisationthe contracted (offered) values of realised works are directlycomparable that is they can be classified based on the totaloffered revalorised value for the realisation of works (Table 3)

Classification of projects according to the total value forthe realisation of works can be of great importance whentrainingtesting of formed models for estimation For thatreason an additional input parameter was introduced which

6 Complexity

Table 5 Outputs from the models

Number Input data description Data type Unit of measure Min Max Mean valueOutput 1 Total offered cost of realisation Numerical RSD 88335301 39542727611 4570530156Output 2 Total offered duration of realisation Numerical day 5 120 asymp38

divides projects into two subsets (values of up to and over40000000 RSD) (Table 4)The borderline was defined basedon the number of projects whose value does not exceed40000000RSDwhich is 70of the total number of analysedprojects whereas the mean value of offered price for all theprojects is also approximately similar to this value (Table 5)

According to the previously defined subject of studytwo outputs from the model were planned the total offeredprice for realisation and total amount of time offered for therealisation of a project

The next step in preparing of the data base is the norma-lisation process Normalisation of data presents the processof reducing of certain data to the same order of magnitudeWhat is achieved in this way is for the data to be analysedwith the same significance when forming a predictionmodelthat is to avoid neglecting of data with a smaller order ofmagnitude range at the very beginning This is the mainreason why the normalisation is necessary that is why itis essential to transform the values into the same range bymoving the range borderlinesThe normalisation process wascarried out for the entire set of 166 analysed projects

Before the normalisation process is applied consideringthe fact that a model based on artificial intelligence will beformed it is necessary to divide the final set of 166 projectsinto a set that will be used for the training of a modelas well as the one used for the testing of formed modelsDefining of which data subset will be used for the trainingof a model and which one for its testing is not entirely basedon the random samplemethodThe comparative analysis wascarried out of the number of projects by categories based onthe offered revalorised total price as well as the offered timefor realisation of all works

When choosing the projects that belong to the trainingsubset particular attention was paid to the fact that the mini-mum and maximum values of all parameters belong to therange of this set In addition all groups of projects shouldbe equally present in both sets according to the offeredvalue and time for realisation Finally the testing subset com-prised 17 pseudorandomly chosen projects (with mentionedrestrictions) whereas the remaining 149 projects constitutethe training subset

The most commonly used forms of data normalisationbeing the simplest ones at the same time are the min-maxnormalisation andZero-Meannormalisation [61] whichwere applied in the conducted research

The first step in forming ofmodels for estimation by usingANNs relates to defining of the number of hidden layersAccording to Huang and Lipmann [62] there is no need touse ANNs with more than two hidden layers which has beenconfirmed by many theoretical results and a number of sim-ulations in various engineering fields Moreover according

to Kecman [63] it is recommendable to start the solving ofa problem by using a model with one hidden layer

By choosing the optimal number of neurons it is neces-sary to avoid two extreme cases omission of basic functions(insufficient number of hidden neurons) and overfitting (toomany hidden neurons) In order to achieve proper general-isation ldquopowerrdquo of an ANN model it is necessary to applythe cross-validation procedure owing it to the fact that goodresults during the training process do not guarantee propergeneralisation ldquopowerrdquoWhat ismeant by generalisation is theldquoabilityrdquo of an ANN model to provide satisfactory results byusing data which were not known to the model during thetraining (validation subset)

For the purpose of the cross-validation procedure withinthe training subset 17 pseudorandomly chosen projects weretaken based on the same principle as within the testingsubset that is the equal percentage share of projects in termsof value If there is not too much difference that is deviationbetween estimated and expected values percentage error(PE) or absolute percentage error (APE) or mean absolutepercentage error (MAPE) in all three subsets (trainingvalidation and testing) it can be considered that this is theactual generalisation power of the formed ANN model thatis that there is no ldquooverfittingrdquo

All the models for estimation of costs and duration wereformed in the Statistica 12 software package in which it ispossible to define two types of ANN models MLP (Multi-layer Perceptron) and RBF (Radial Basis Function) modelsAccording to Matignon [64] both models are used to dealwith classification problems while the MLP models are usedto deal with regression problems and RBF models with clus-tering problems Since the subject of the research involves theestimation of costs and duration that is belonging to regres-sion problems only the MLP ANNmodels were formed

Activation function of output neurons is mainly linearwhen it comes to regression problems When the activationfunctions of hidden neurons are concerned the most com-monly used functions are logistic unipolar and sigmoidalbipolar (hyperbolic tangent being the most commonly usedone) [63] In accordance with this recommendation for all themodels activation functions for hidden neurons logistic sig-moid and hyperbolic tangent were used while the activationfunction identity was used for output neurons (Table 6)


Based on previously defined inputs outputs and definedparameters in each iteration 10000 ANNMLP models wereformed and onemodelwith the smallest estimation errorwaschosenThe number of input and output neurons was definedby the number of inputs and outputs whereas the number of

Complexity 7

Table 6 Activation functions of MLP ANNmodels

Function Expression Explanation RangeIdentity a Activation of neurons is directly forwarded as output (minusinfin +infin)Logistic sigmoid

11 + eminusa ldquoSrdquo curve (0 1)

Hyperbolic tangent ea minus eminusaea + eminusa

Sigmoid curve similar to logistic function but featuring betterperformances because of the symmetry it has Ideal for MLP ANN

models especially for hidden neurons(minus1 +1)

Table 7 ANN models (min-max)

Model NetworkActivation

function hiddenneurons

Activationfunction output

neurons

MAPE training(cost) []

MAPE training(duration) []

MAPE testing(cost) []

MAPE testing(duration) []

ANN 1 MLP 12-4-2 Tanh Identity 4279 3185 4054 3548ANN 2 MLP 11-6-2 Logistic Identity 4188 3182 2697 3022ANN 3 MLP 12-6-1 Tanh Identity 3302 2538 ANN 4 MLP 11-7-1 Tanh Identity 3916 2688 ANN 5 MLP 12-8-1 Tanh Identity 3416 2626ANN 6 MLP 11-10-1 Logistic Identity 3329 3516

Table 8 ANN models (Zero-Mean)




neurons





ANN 7 MLP 12-7-2 Logistic Identity 5216 3089 3796 3423ANN 8 MLP 11-8-2 Logistic Identity 4804 3206 4254 3420ANN 9 MLP 12-4-1 Tanh Identity 3749 2022 ANN 10 MLP 11-8-1 Tanh Identity 4099 2828 ANN 11 MLP 12-8-1 Tanh Identity 3232 3720ANN 12 MLP 11-5-1 Tanh Identity 3313 3559

hidden neurons was limited to maximum of 10 A total of 12ANN models were chosen 6 of them being normalised bythe min-max procedure and the remaining 6 by the 119885 Scoreprocedure

By using the ANN 1 and ANN 2 models simultaneousestimation of costs and duration was carried out The ANN1 model had 12 inputs and 2 outputs whereas the ANN 2model was formed by using 11 inputs Elimination of oneinput parameter followed the analysis of the influence of inputparameters which showed that the realisation zone (11i) hasthe smallest influence on the values of the output data Thesame principle was applied in forming of ANN 3 and ANN 4models for the estimation of costs only as well as the ANN 5and ANN 6 models for the estimation of time needed for therealisation of works

Table 7 shows the chosen models (min-max normali-sation) with defined characteristics of models and definedaccuracy of estimation expressed through the MAPE

Forming of the remaining 6 ANN models with datawhose normalisation was performed by using the Zero-Meannormalisation was carried out in the same way In this case aswell the realisation zone (11i) has the smallest impact on the

output values Table 8 provides an outline of chosen modelswith defined characteristics of models and defined accuracyof estimation expressed through the MAPE

The comparative analysis of presented models clearlyshows that the greater accuracy of estimation is achievedby models formed based on the data prepared by the min-max normalisation procedure The accuracy of estimation offormed models is unsatisfactory that is being considerablylarger than the desirable plusmn15 for the costs of construction

The first step in the forming of models for estimation byusing the SVM as well as the ANNmodels relates to definingof input and output data In the process of forming of SVMmodels the used data had previously been prepared by apply-ing the min-max normalisation as it was proved that using itresults in the greater accuracy in the case of ANN modelsMoreover only the models for separate estimation of costsand duration of construction were formed The main reasonfor this lies in the fact that the greater accuracy is achievedby separate estimation that is by forming of separatemodelswhichwas proved on theANNmodelsHowever the softwarepackage Statistica 12 itself does not provide the option ofsimultaneous estimation of several parameters by using the

8 Complexity

Table 9 Functions of error of SVMmodels

SVM type Error function Minimize subject to

Type 1 12119908119879119908 + 119862119873sum119894=1

120585119894+ 119862 119873sum119894=1

120585lowast119894

119908119879120601(119909119894) + 119887 minus 119910

119894le 120576 + 120585lowast

119894119910119894minus 119908119879120601(119909

119894) minus 119887119894le 120576 + 120585

119894120585119894 120585lowast119894ge 0 119894 = 1119873

Type 2 12119908119879119908 minus 119862(]120576 + 1119873119873sum119894=1

(120585119894+ 120585lowast119894)) (119908119879120601(119909

119894) + 119887) minus 119910

119894le 120576 + 120585lowast

119894119910119894minus (119908119879120601(119909

119894) + 119887119894) le 120576 + 120585

119894120585119894 120585lowast119894ge 0 119894 = 1119873 120576 ge 0

Table 10 SVMmodels (min-max)

Model 119862 120576 121205902 MAPE training(cost) []




SVM 1 20 0001 0083 2528 1547 SVM 2 20 0001 0091 2396 706 SVM 3 20 0001 0083 2921 2459SVM 4 20 0001 0091 3075 2277

SVMWithin thementioned software package two functionsof error in the forming of SVMmodels are offered (Table 9)

For Type 1 (epsilon-SVM regression) it is necessary todefine parameter capacity (C) and epsilon (120576) insensitivityzone At the same time for Type 2 (nu-SVM regression) itis necessary to define parameter capacity (119862) and Nu (120574)The value of parameters C and 120576 ranges between 0 and infinwhereas the value of parameter 120574 ranges between 0 and 1 It isalso necessary to choose one of the offered Kernel functionslinear polynomial RBF or sigmoid RBF kernel functionpresents themost frequently used kernel function for formingof SVMmodels

K (xi xj) = exp (minus 1212059021003817100381710038171003817x minus xi10038171003817100381710038172) 120590 ndash width of RBF function (1)

When using the RBF kernel function it is necessary to definethe parameter 120574 = 121205902 For the purpose of forming of themodel the function of error Type 1 was used

A total of four SVM models were formed with the samenumber of input parameters as in the case of ANN modelsfrom the perspective of input parameters SVM 1 = ANN3 and ANN 9 SVM 2 = ANN 4 and ANN 10 SVM 3 =ANN 5 and ANN 11 and SVM 4 = ANN 6 and ANN 12The reason for this is the easier comparative analysis ofresults obtained by using the listed models Table 10 showscharacteristics of formed models with defined accuracy ofestimation expressed through the MAPE

After forming the SVM models it is evident that theyprovide greater accuracy of estimation of both costs andduration of projects aswellThe shown accuracy of estimationmade through the MAPE is not sufficient for the choiceof a model but it is necessary to carry out the analysis ofaccuracy of estimation for each of the projects separatelyespecially those from the testing subset The estimation erroris expressed through the PE (percentage error) which isshown inTable 11 for the estimation of costs and inTable 12 forthe estimation of duration of projects Defining of the PE is of

particular importance for the estimation of costs in order tocarry out the comparative analysis with the desired accuracyofplusmn15 for each of the projects from the testing set separately

Errors in the estimation of duration are higher whencompared to estimation of offered prices for the realisation ofworks since the investor defined in the tender documenta-tion the maximum possible duration of works which is morethan an optimistic prediction In other words the offeredtime is not the result of the estimation by the contractor butthe limitation set by the investor For this reason contractorsadopted automatically the maximum offered duration ofworks

Additionally it can be stated that models for separateestimation of costs and duration provide a higher level ofaccuracy than those which carry the estimation out simul-taneously which is specifically the case with ANN modelsThe reason for this lies in the fact that the impact of inputparameters on the output ones is not the same for theestimation of costs and duration of works

Figure 1 shows the influence of input parameters on theoutput ones in the case of estimation of costs for the ANN3 and ANN 4 models The input data are divided into fourcategories that is two groups and two independent piecesof input data The first group presents inputs which relate toamounts of materials needed for the realisation of works onthe roadway construction and landscaping the second onerelates to the share of works by the remaining groups of worksfrom the bill of works whereas the independent input datarelates to the realisation zone and the category of the project

Figure 2 Illustrates the influence of input parameters forthe ANN 5 and ANN 6 models that is models for theestimation of duration of projects

Based on the graphs given above it can clearly be noticedthat the second group of input data has a considerably greaterinfluence on the estimation of project duration diminishingthe influence of the first group Moreover the category ofthe project is of approximately the same significance for the

Complexity 9

Table 11 PE for estimation of costs testing set

Expected cost [RSD] PE for the cost for the model testing []ANN 1 ANN 2 ANN 3 ANN 4 SVM 1 SVM 2

(1) 264822252 minus10264 5014 702 minus2216 3513 2926(2) 374599606 minus13549 2481 minus6412 minus685 1090 295(3) 431645020 minus7175 minus335 minus5863 minus419 330 minus369(4) 440674587 minus884 minus2848 minus2643 4385 minus10106 minus1716(5) 589457764 minus6160 minus5676 minus3047 minus7195 658 minus668(6) 622826297 10817 6110 5849 7007 minus036 1147(7) 795953114 minus824 minus2217 minus2671 minus4593 minus2182 minus1467(8) 1440212926 minus2513 minus1058 1293 2789 minus3166 381(9) 1549908198 minus3301 minus1510 minus2531 minus2305 685 693(10) 2429815868 minus1178 5589 minus1172 minus509 minus695 246(11) 2929337643 minus623 minus186 minus224 minus1882 025 019(12) 3133158369 minus2266 minus1792 minus952 minus2311 minus658 minus387(13) 4862894636 minus1212 minus3733 469 609 minus472 268(14) 6842852313 minus6168 minus2573 minus5810 minus5168 minus1391 minus686(15) 7482821100 816 2596 708 465 795 428(16) 12197147998 697 1185 2616 3090 268 190(17) 26733389415 minus463 minus951 minus191 minus062 minus231 minus121

MAPE 4054 2697 2538 2688 1547 706

Table 12 PE for estimation of duration testing set

Expected duration [day] PE for the duration of the model testing subset []ANN 1 ANN 2 ANN 5 ANN 6 SVM 3 SVM 4

(1) 12 minus2688 minus2747 minus2084 minus1789 minus2945 minus6918(2) 26 4900 1441 1484 3766 631 minus246(3) 20 1124 minus421 770 2481 minus403 minus753(4) 35 minus063 107 913 minus4502 1106 2842(5) 34 1986 2268 1373 2995 3581 3624(6) 17 3471 2611 minus1314 minus1266 minus580 001(7) 17 minus3919 minus2734 minus2132 3501 3142 2706(8) 20 minus8638 minus3859 minus5730 minus9300 minus8395 minus4162(9) 25 390 minus1467 minus353 minus219 542 minus093(10) 35 minus1614 1357 minus055 617 minus2822 minus172(11) 25 minus6140 minus6219 minus5517 minus5497 minus1728 minus2371(12) 38 066 minus1499 085 minus174 2163 1926(13) 41 minus12467 minus12300 minus11363 minus10376 minus6612 minus6311(14) 60 minus5147 minus5354 minus4783 minus5672 minus2393 minus2482(15) 60 4208 4397 3767 4403 1742 1395(16) 60 minus2936 minus2488 minus1708 minus3075 minus2226 minus1972(17) 75 561 109 1214 129 795 729

MAPE 3548 3022 2626 3516 2459 2277

estimation of both costs and duration as well whereas therealisation zone has a considerably greater impact on theestimation of duration of the project

5 Conclusions

Based on the results presented above a conclusion wasdrawn that a greater accuracy level in estimating of costs and

duration of construction is achieved by using of models forseparate estimation of costs and durationThe reason for thislies primarily in the different influence of input parameterson the estimation of costs in comparison with the estimationof duration of the project By integrating them into a singlemodel a compromise in terms of the significance of input datais made resulting in the lower precision of estimation whenit comes to ANNmodels

10 Complexity

964

903

3537834

571

508

320

259

263

266341

1234

Amount of crushed stone

Amount of curbs

Amount of asphalt base layer

Amount of asphalt surface layer

Amount concrete prefabricated elements

Preparation works

Earthworks

Drainage works

Trac signalisation works

Other works

Works realisation zone

Project category

1071

697

3424

1209

628

538

366

238

241

280

1308


Amount of curbs




Preparation works

Earthworks

Drainage works


Other works

Project category

Figure 1 Analysis of sensitivity (ANN 3 and ANN 4) estimation of construction costs

788713736

810842

698780

750971

916706

1290

Amount of crushed stoneAmount of curbs

Amount of asphalt base layerAmount of asphalt surface layer

Amount concrete prefabricated elementsPreparation works

EarthworksDrainage works

Trac signalisation worksOther works

Works realisation zoneProject category

739

854

600

787

1021

826

823

813

1156

792

1589


Amount of curbs




Preparation works

Earthworks

Drainage works


Other works

Project category

Figure 2 Analysis of sensitivity (ANN 5 and ANN 6) estimation of duration of construction

SVM models feature a greater capacity of generalisationproviding at the same time greater accuracy of estimationfor the estimation of both costs and duration of projects aswell In both cases the greatest accuracy of estimation isachieved by the SVMmodel with 11 input parameters that iswithout a more considerable influence of the project realisa-tion zone which implies that the investors did not pay parti-cular attention when defining the offered price and time towhether the works will be realised within the city zone or thesuburbs

Extending of the data base in terms of the very subjectof a contract that is the future construction object and theintroduction of parameters such as the length of the sectionthe roadway width the urban road category (boulevard sidealley etc) the length of cycling lanes areas intended forparking pedestrian lanes and plateaus would widely extendthe possibility of application of estimation models There is awide range of potential parameters which can be introducedas the feature of the construction object presenting at thesame time the input data for the estimation model In thisway the possibility of estimating the costs and duration ofconstruction in the contracting phase not only when the billof works is defined (query by tender) but also when the

investor defines only the guidelines that is the conditionsthat the future construction object has to meet (query byfunctional parameters of a future object) is created Formingof a model with functional features of a future constructionobject as input parameters could have a double applicationfrom both perspectives of the investor and the contractorFrom the perspective of the investor if the accuracy ofestimation similar to that of already formed models wasachieved the precision in estimation in the initial phase anddefining of criteria which the future object is to meet wouldbe considerably above the one required by the literature(plusmn50) [10]

The research was conducted for the estimation of costsand duration of realisation for ldquobuildrdquo contracts Thisapproach provides an option to estimate ldquodesign-buildrdquo con-tracts as well that is the experience gained through ldquobuildrdquocontracts can be used in future estimations of ldquodesign-buildrdquocontacts for both the needs of the investor and the contractoras well Application of future models largely depends onthe available information at the time of estimation For thisreason the input data should be adjusted to the availableinformation at the given moment for both the investor andthe contractor as well

Complexity 11



Acknowledgments

The work reported in this paper is a part of the investigationwithin the research project ldquoUtilization of By-Products andRecycled Waste Materials in Concrete Composites in TheScope of Sustainable Construction Development in SerbiaInvestigation and Environmental Assessment of PossibleApplicationsrdquo supported by the Ministry of Education Sci-ence and Technological Development Republic of Serbia (TR36017) This support is gratefully acknowledged The workreported in this paper is also a part of the investigationwithin the research project ldquoOptimization of Architecturaland Urban Planning and Design in Function of SustainableDevelopment in Serbiardquo supported by the Ministry of Edu-cation Science and Technological Development Republic ofSerbia (TR 36042) This support is gratefully acknowledged

References

[1] J Gunner and M Skitmore ldquoComparative analysis of pre-bidforecasting of building prices based on Singapore datardquo Con-struction Management and Economics vol 17 no 5 pp 635ndash646 1999

[2] NMorrison ldquoThe accuracy of quantity surveyorsrsquo cost estimat-ingrdquo Construction Management and Economics vol 2 no 1 pp57ndash75 1984

[3] M Skitmore and H P Lo ldquoA method for identifying high out-liers in construction contract auctionsrdquo Engineering Construc-tion and Architectural Management vol 9 no 2 pp 90ndash1302002

[4] M A Azman Z Abdul-Samad and S Ismail ldquoThe accuracy ofpreliminary cost estimates in PublicWorksDepartment (PWD)of Peninsular Malaysiardquo International Journal of Project Man-agement vol 31 no 7 pp 994ndash1005 2013

[5] A A Aibinu and T Pasco ldquoThe accuracy of pre-tender buildingcost estimates in AustraliardquoConstructionManagement and Eco-nomics vol 26 no 12 pp 1257ndash1269 2008

[6] SM Abou Rizk GM Babey andG Karumanasseri ldquoEstimat-ing the cost of capital projects An empirical study of accuracylevels for municipal government projectsrdquo Canadian Journal ofCivil Engineering vol 29 no 5 pp 653ndash661 2002

[7] J S Shane K R Molenaar S Anderson and C SchexnayderldquoConstruction project cost escalation factorsrdquo Journal of Man-agement in Engineering vol 25 no 4 pp 221ndash229 2009

[8] M Skitmore ldquoRaftery curve construction for tender price fore-castsrdquo Construction Management and Economics vol 20 no 1pp 83ndash89 2002

[9] B Ivkovic and Z Popovic Upravljanje projektima u građev-inarstvu Građevinska knjiga Beograd 2005

[10] A Ashworth Cost Studies of Buildings Pearson EducationLimited England UK 5th edition 2010

[11] M Asif and R M W Horner Economical Construction DesignUsing Simple Cost Models University of Dundee 1989

[12] R M Horner K J McKay and M M Saket ldquoCost-significancein Estimating and Control Symposium Organization in Con-structionrdquo in Proceedings of the Cost-significance in Estimating

and Control Symposium Organization in Construction pp 571ndash585 Opatija Croatia 1986

[13] H Adeli ldquoNeural networks in civil engineering 1989ndash2000rdquoComputer-AidedCivil and Infrastructure Engineering vol 16 no2 pp 126ndash142 2001

[14] A M Elazouni I A Nosair Y A Mohieldin and A GMohamed ldquoEstimating resource requirements at conceptualdesign stage using neural networksrdquo Journal of Computing inCivil Engineering vol 11 no 4 pp 217ndash223 1997

[15] V Albino and A C Garavelli ldquoA neural network application tosubcontractor rating in construction firmsrdquo International Jour-nal of Project Management vol 16 no 1 pp 9ndash14 1998

[16] A W T Leung C M Tam and D K Liu ldquoComparative studyof artificial neural networks and multiple regression analysisfor predicting hoisting times of tower cranesrdquo Building andEnvironment vol 36 no 4 pp 457ndash467 2001

[17] S Rebano-Edwards ldquoModelling perceptions of building qual-ity-A neural network approachrdquo Building and Environment vol42 no 7 pp 2762ndash2777 2007

[18] D Vouk DMalus and I Halkijevic ldquoNeural networks in econ-omic analyses of wastewater systemsrdquo Expert Systems withApplications vol 38 no 8 pp 10031ndash10035 2011

[19] J-H Chen and S C Hsu ldquoHybrid ANN-CBR model for dis-puted change orders in construction projectsrdquo Automation inConstruction vol 17 no 1 pp 56ndash64 2007

[20] N B Chaphalkar K C Iyer and S K Patil ldquoPrediction ofoutcome of construction dispute claims using multilayer per-ceptron neural network modelrdquo International Journal of ProjectManagement vol 33 no 8 article no 1808 pp 1827ndash1835 2015

[21] H P Tserng G-F Lin L K Tsai and P-C Chen ldquoAn enforcedsupport vector machine model for construction contractordefault predictionrdquo Automation in Construction vol 20 no 8pp 1242ndash1249 2011

[22] K C Lam E Palaneeswaran and C-Y Yu ldquoA support vectormachine model for contractor prequalificationrdquo Automation inConstruction vol 18 no 3 pp 321ndash329 2009

[23] M Y Cheng and A F V Roy ldquoEvolutionary fuzzy decisionmodel for cash flow prediction using time-dependent supportvector machinesrdquo International Journal of Project Managementvol 29 no 1 pp 56ndash65 2011

[24] M Wauters and M Vanhoucke ldquoSupport Vector MachineRegression for project control forecastingrdquo Automation inConstruction vol 47 pp 92ndash106 2014

[25] V Mucenski M Trivunic G Cirovic I Pesko and J DrazicldquoEstimation of recycling capacity of multistorey building struc-tures using artificial neural networksrdquo in Acta PolytechnicaHungarica vol 10 pp 175ndash192 4 edition 2013

[26] S O Cheung P S P Wong A S Y Fung and W V CoffeyldquoPredicting project performance through neural networksrdquoInternational Journal of Project Management vol 24 no 3 pp207ndash215 2006

[27] H M Gunaydin and S Z Dogan ldquoA neural network approachfor early cost estimation of structural systems of buildingsrdquoInternational Journal of Project Management vol 22 no 7 pp595ndash602 2004

[28] M Arafa andM Alqedra ldquoEarly stage cost estimation of build-ings construction projects using artificial neural networksrdquoJournal of Artificial Intelligence Research vol 4 no 1 pp 63ndash752011

[29] M Bouabaz and M Hamami ldquoA cost estimation model forrepair bridges based on artificial neural networkrdquo AmericanJournal of Applied Sciences vol 5 no 4 pp 334ndash339 2008

12 Complexity

[30] D P Alex M Al Hussein A Bouferguene and S FernandoldquoArtificial neural network model for cost estimation City ofedmontonrsquos water and sewer installation servicesrdquo Journal ofConstruction Engineering and Management vol 136 no 7 pp745ndash756 2010

[31] I SiqueiraNeural Network-Based Cost Estimating [Master the-sis] University of Montreal Quebec Department of BuildingCivil and Environmental Engineering Canada 1999

[32] S-H An U-Y Park K-I Kang M-Y Cho and H-H CholdquoApplication of support vectormachines in assessing conceptualcost estimatesrdquo Journal of Computing in Civil Engineering vol21 no 4 pp 259ndash264 2007

[33] H Adeli and M Wu ldquoRegularization neural network for con-struction cost estimationrdquo Journal of Construction Engineeringand Management vol 124 no 1 pp 18ndash24 1998


[35] B Vahdani S M Mousavi M Mousakhani M Sharifi andH Hashemi ldquoA neural network modela based on supportVector machine for conceptual cost estimation in constructionprojectsrdquo Journal of Optimization in Industrial Engineering vol10 p 11 2012

[36] F Kong X-J Wu and L-Y Cai ldquoA novel approach based onsupport vector machine to forecasting the construction projectcostrdquo in Proceedings of the 2008 International Symposium onComputational Intelligence and Design (ISCID rsquo08) pp 21ndash24China October 2008

[37] F Kong XWu andLCai ldquoApplication of RS-SVM in construc-tion project cost forecastingrdquo in Proceedings of the 4th Interna-tional Conference onWireless Communications Networking andMobile Computing (WiCOM rsquo08) Dalian China October 2008

[38] G Kim J Shin S Kim and Y Shin ldquoComparison ofSchool Building ConstructionCosts EstimationMethodsUsingRegression Analysis Neural Network and Support VectorMachinerdquo Journal of Building Construction and Planning Re-search vol 01 no 01 pp 1ndash7 2013

[39] G H Kim S H An and K I Kang ldquoComparison of construc-tion cost estimatingmodels based on regression analysis neuralnetworks and case-based reasoningrdquo Building and Environ-ment vol 39 no 10 pp 1235ndash1242 2004

[40] R Sonmez ldquoConceptual cost estimation of building projectswith regression analysis and neural networksrdquo Canadian Jour-nal of Civil Engineering vol 31 no 4 pp 677ndash683 2004

[41] G H Kim D S Seo and K I Kang ldquoHybrid models of neuralnetworks and genetic algorithms for predicting preliminarycost estimatesrdquo Journal of Computing in Civil Engineering vol19 no 2 pp 208ndash211 2005

[42] W F Feng W J Zhu and Y G Zhou ldquoThe application of gen-etic algorithm and neural network in construction cost esti-materdquo in Proceedings of the Third International Symposium onEletronic Commerce and SecurityWorkshop (ISECS rsquo10) pp 151ndash155 Guangzhou China 2010

[43] M Y Cheng and C J Huang ldquoConstruction ConceptualCost Estimates Using Evolutionary Fuzzy Neural InferenceModelrdquo in Proceedings of the 20th International Symposium onAutomation and Robotics in Construction pp 595ndash600 Eind-hoven The Netherlands September 2003

[44] M-Y Cheng H-C Tsai and E Sudjono ldquoConceptual costestimates using evolutionary fuzzy hybrid neural network forprojects in construction industryrdquo Expert Systems with Applica-tions vol 37 no 6 pp 4224ndash4231 2010

[45] M Cheng andYWu ldquoConstructionConceptual Cost EstimatesUsing Support Vector Machinerdquo in Proceedings of the 22ndInternational Symposium on Automation and Robotics in Con-struction pp 1ndash5 Ferrara Italy September 2005

[46] S G Deng and T H Yeh ldquoUsing least squares support vectormachine to the product cost estimationrdquo vol 25 no 1 pp 1ndash162011

[47] M Y Cheng H S Peng Y W Wu and T L Chen ldquoEstimateat completion for construction projects using evolutionarysupport vector machine inference modelrdquo Automation in Con-struction vol 19 no 5 pp 619ndash629 2010

[48] X-Z Wang X-C Duan and J-Y Liu ldquoApplication of neuralnetwork in the cost estimation of highway engineeringrdquo Journalof Computers vol 5 no 11 pp 1762ndash1766 2010

[49] H Al-Tabtabai A P Alex and M Tantash ldquoPreliminary costestimation of highway construction using neural networksrdquoCost Engineering vol 41 no 3 pp 19ndash24 1999

[50] T Hegazy and A Ayed ldquoNeural network model for parametriccost estimation of highway projectsrdquo Journal of ConstructionEngineering and Management vol 124 no 3 pp 210ndash218 1998

[51] J Sodikov ldquoCost estimation of highway projects in developingcountries artificial neural network approachrdquo Journal of theEastern Asia Society for Transportation Studies vol 6 pp 1036ndash1047 2005

[52] N I El-Sawahli ldquoSupport vector machine cost estimationmodel for road projectsrdquo Journal of Civil Engineering and Archi-tecture vol 9 pp 1115ndash1125 2015

[53] A AttalDevelopment of neural network models for prediction ofhighway construction cost and project duration [Master thesis]Department of Civil Engineering and the Russ College of Engi-neering and Technology Ohio University Ohio Ohio USA2010

[54] S Bhokha and S O Ogunlana ldquoApplication of artificial neuralnetwork to forecast construction duration of buildings at thepredesign stagerdquo Engineering Construction and ArchitecturalManagement vol 6 no 2 pp 133ndash144 1999

[55] B Hola and K Schabowicz ldquoEstimation of earthworks execu-tion time cost by means of artificial neural networksrdquo Automa-tion in Construction vol 19 no 5 pp 570ndash579 2010

[56] Y-RWang C-Y Yu andH-H Chan ldquoPredicting constructioncost and schedule success using artificial neural networksensemble and support vector machines classification modelsrdquoInternational Journal of Project Management vol 30 no 4 pp470ndash478 2012

[57] I Pesko G Cirovic V Mucenski and Z Tepic ldquoAnalysis andpreparation of date in neural networks calculation stage forthe purpose of creating business proposalsrdquo in Proceedings ofthe 10th International Conference Organization Technology andManagement in Construction Book of Abstracts p 57 SibenikCroatia September 2011

[58] I Pesko M Trivunic G Cirovic and V Mucenski ldquoA prelimi-nary estimate of time and cost in urban road construction usingneural networksrdquo Tehnicki vjesnik vol 3 no 20 pp 563ndash5702013

[59] R D Stewart R M Wyskida and J D Johannes Cost Esti-matorrsquos Reference Manual John Wiley amp Sons USA 1995

[60] D S Remer andH R Buchanan ldquoEstimating the cost for doinga cost estimaterdquo International Journal of Production Economicsvol 66 no 2 pp 101ndash104 2000

[61] F J M Lopez S M Puertas and J A T Arriaza ldquoTraining ofsupport vector machine with the use of multivariate normaliza-tionrdquo Applied Soft Computing vol 24 pp 1105ndash1111 2014

Complexity 13

[62] W Y Huang and R P Lipmann Neural Net and TraditionalClassifiers uNeural Information Processing Systems D Z Ander-son Ed American Institute of Physics New York NY USA1988

[63] V Kecman Learning and Soft Computing Support Vector Ma-chines Neural Networks and Fuzzy Logic Models MassachusettsInstitute of Technology (MIT)MassachusettsMassUSA 2001

[64] R Matignon Neural Networks Modeling using SAS EnterpriseMIner ISBN 1-4184-2341-6 2005


Complexity





Editorial Board




Contents










and Andrej Soltesz5























2 Complexity










1 Introduction



















2 Complexity










3Complexity





















Z =wiZi =sumwiZi

sumwi= w1w1 +w2

Z1 +w2

w1 +w2Z1


1



N

M

Bel 3

el 2

el 1A

h = 3 m

b

d2

d2

y Tf = 20degC


4 Complexity








Inputlayer

Hiddenlayer 1

Hiddenlayer 2

Hiddenlayer 3

Hiddenlayer 4

Outputlayer

X

Y

Premiseparameters


A1

w1

w2

w1

w2

w2Z2

w1Z1

x1 x2

x1 x2

A2

B1

B2

N

N

120587

120587

f

f

sumZ


5Complexity











20

15

10Out

put

5

00 50 100 150 200



6 Complexity




10

8

Out

put

4

6

2

00 5 10 15 20

Data set index25 30


Testing data ()20

15

10

Out

put

5

00 10 20 30 40






7Complexity











20

15

Out

put

5

10

00 50 100 150 200 250

Index300 350


Testing data 20

15

Out

put

0

5

10

minus50 10 20 30 40 50 60

Index70 80

FIS output


8 Complexity




5 Conclusion





1580minus2006

1

2

3

4

5

6

7

8

30

Input[4040310503025]

Plot points101

Moveleft

Help Close

right down up


50 30 50 2 4 06 15 01 05 050


9Complexity


20

15

Out

put

5

10

00 5 10 15 20

Index25 30



Checking data

Columndimensions







3000 3000 200 100 010 02 588 553

3000 3000 200 100 030 03 214 187

3000 3000 300 100 010 05 294 344

3000 3000 300 100 040 04 132 132

3000 3000 400 100 040 01 208 183

4000 4000 200 100 010 02 776 798

4000 4000 200 100 040 0 382 383

4000 4000 300 100 020 04 356 374

4000 4000 300 100 030 05 216 228

4000 4000 300 100 040 0 38 377

4000 4000 400 100 010 01 1014 99

4000 4000 400 100 030 03 344 361

4000 4000 300 060 020 02 556 552

4000 4000 300 150 010 0 12 117

4000 4000 300 150 050 04 138 128

5000 5000 200 100 010 03 983 101

5000 5000 300 100 010 05 528 559

5000 5000 400 100 030 04 384 361

5000 5000 200 060 020 01 95 903

5000 5000 200 060 050 0 318 337

5000 5000 400 060 030 02 57 553

5000 5000 400 060 050 0 352 353

5000 5000 400 060 050 01 314 313

5000 5000 400 060 050 02 28 281

5000 5000 400 060 050 03 232 252

5000 5000 400 060 050 04 188 22

5000 5000 400 060 050 05 148 186

10 Complexity





Data Availability




References





















11Complexity


























12 Complexity









1 Introduction




















2 Methodology


2 Complexity









Envi

ronm

ent

Database Modelbase

Strategicmanagement

level



Dialog

Users

41

2 3


3Complexity












350 mm

Asphalt


Subgrade


4 Complexity






















2

Asphalt layer


3 Thickness (3 6 9 12 15 18 and 21 cm)


6







Input data x1

Input data xm

Inputlayer

Outputlayer

Hidden layers

Output data

⋮

⋮ ⋮


5Complexity













6 Complexity









A B C

Figure 5 Test cases

09940992


0965 0994

0998

0996

0997

0999


00

1

2

3

Pred

icte

d str

ess v

alue

(MPa

)

4

5

6

7



R2 = 0993

Case AEquality line


4 5 6 7


7Complexity





R2 = 09652

Pred

icte

d str

ess v

alue

s (M

Pa)

Case B

0

1

2

3

4

minus1

Equality line

0 1 2 3 4








(cm)


(MPa)


C2Modulus of



Stress 05 20ndash100 10 1400 and 9700




2



0

1

2

Stre

ss (M

Pa)

3

4


14 16 18 20


8 Complexity




4 Conclusions



Data Availability




References



0

1

2

Stre

ss (M

Pa)





0

1

2

Stre

ss (M

Pa)




05

15

25




Stre

ss (M

Pa)


9Complexity


























10 Complexity









1 Introduction











2 Complexity












119891 119883 997888rarr 119884 (1)



Complexity 3






Yes

No











surface







4 Complexity




(villages)



















Complexity 5


















119910 = 119891 (1199091 1199092 1199093 1199094 1199095 1199096 1199097) + 120576 (2)


119910 = 119891 (1199091 1199092 1199093 1199094 1199095 1199096 1199097) (3)


6 Complexity


119901 119910 1199091 1199092 1199093 1199094 1199095 1199096 11990975 5658 968 09 06 00 1968 6025 0013 1359 3292 09 05 6000 21420 00 0023 4276 1860 09 03 11160 780 375 1000037 4895 1860 01 03 2400 1000 00 0046 3230 800 09 05 1920 1818 1397 0059 1813 800 01 03 720 00 00 400067 19723 4131 09 05 3960 10960 2071 2586482 1611 1650 01 01 3000 00 00 0094 2500 1470 01 02 3445 939 385 00101 8005 1104 09 08 2959 14691 933 3746




1199091 133379 27500 560000 788491199092 059 010 090 0271199093 049 010 090 0161199094 30026 000 221200 345271199095 19316 000 214200 325841199096 10524 000 60250 121131199097 32429 000 300000 60326



119877 (119910 119910) = cov (119910 119910)120590푦120590푦


(4)









119867 asymp radic119873119872 (5)


NNP = NNW + NNB lt 119871 (6)


Complexity 7



119871 0901 0240 821 1455 1117119881 0879 0228 810 1088 578119879 0881 0233 805 1269 833

minus100

minus080

minus060

minus040

minus020

000

020

040

060

080

100


RT(y y)

RL(y y)

(a)

0

100

200

300

400

500

0 100 200 300 400 500

２－

３T

２－３L

(b)












8 Complexity

000

020

040

060

080

100

000 020 040 060 080 100

minus100

minus080

minus060

minus040


RV(y y)

RL(y y)

(a)

0

100

200

300

400

500

0 100 200 300 400 500

２－

３V

２－３L

(b)








SSE = sum푝isin퐿

(119910푝 minus 119910푝)2 (7)





Complexity 9


ANN 119877 (119910 119910) RMSE119871 119881 119879 119871 119881 119879MLPe-l 7-2-1






MLPs-l 7-5-1 119871 119881 119879MAPE


PEmax




100381610038161003816100381610038161003816100381610038161003816

PE푝 = 100381610038161003816100381610038161003816100381610038161003816119910푝 minus 119910푝119910푝

100381610038161003816100381610038161003816100381610038161003816 100PEmax = max (PE푝)

(8)



10 Complexity

0

500

1000

1500

2000

2500

3000

0 500 1000 1500 2000 2500 3000y




y













Acknowledgments


References








Complexity 11































1 Introduction





2 Complexity










Complexity 3











4 Complexity






119908(119897)new119894119895 = 119908(119897)old119894119895 minus 120578

120597120576119896

120597119908(119897)119894119895 (1)


119891 (119909) =1

1 + 119890minus119909 (2)


Δ119908(119897)119894119895 (119905) = minus120578120597120576119896

120597119908(119897)119894119895+ 120572Δ119908(119897)119894119895 (119905 minus 1) 0 lt 119905 lt 1 (3)





2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2005

Year

100000

300000

500000

700000

900000

1100000

Pric

e






Complexity 5

Table1Ba

sicinpu

tsinto

them

odel

theird

efinitio

nandbasic

characteris

tics

Varia

ble

Varia

bled

efinitio

n(according

tothes

ources


Minim

umMaxim

umMean

Standard

deviation

Unito

fmeasure

GDPlowast

GDP(grossdo

mestic

prod

uct)reflectsthe

totalvalue

ofallgoo

dsandservices

prod

uced

lessthe

valueo

fgoo

dsandservices

used


ptionin

theirp

rodu

ction

5142

3032820

47776356

7206493

5mileuro

BDPperc

apitalowast

GDPperc

apita

atmarketp

rices

iscalculated

asther

atio

ofGDPatmarketp

rices

tothea

verage

popu

latio

nof

aspecific

year

4200

84400

2382963

1575946

euro

RealGDPgrow

thratelowast

Thec

alculationof

thea

nnualgrowth

rateof


allowcomparis

onso

fthe

dynamicso

fecono

micdevelopm

entb

othover

timea

ndbetweenecon

omieso


119

163

384

Inequalityof

income

distr

ibutionlowast

Ther

atio

oftotalincom

ereceivedby

the2

0of

thep

opulationwith

theh

ighestincome(top

quintile)to

thatreceived

bythe2

0of

thep

opulationwith

thelow

estincom

e(lowestq

uintile)

32

83

483

116

-

Totalu


Unemploymentrates

representu

nemployed

person

sasa

percentage

ofthelabou

rforce

340

2750

907

432

Averagea

nnualn

etearningslowast

Netearnings

arec

alculated

from

grosse

arning

sbydedu

ctingthee


contrib

utions

andincometaxes

andadding

family

allowancesinthec

aseo

fhou

seho

ldsw

ithchild

ren

155077

3849018

1717581

1044515

euro

FDIlowastlowast


ityflo

wsinther

eportin

gecon

omyItis

thes

umof


ento

fearning

sandotherc

apita

lminus29679

734010

27948

67501

mil$

HICP-inflatio

nratelowast

Harmon

isedIndiceso

fCon

sumer

Prices

(HICPs)a

redesig

nedforinternatio

nalcom

paris

onso

fconsum

erpriceinfl

ation

minus16

153

238

220

VAT(

)lowastlowastlowast

Thev

alue


VATin

theE

urop

eanUnion

(EU)isa

generalbroadlybased

consum

ptiontaxassessed

onthev

alue

addedto

good

sand

services

1527

205

257

Taxeso

nprop

ertyas

of


Taxon

prop

ertyisdefin

edas

recurrentand

nonrecurrent

taxeso

ntheu

seownershipor

transfe

rof

prop

ertyTh

eseinclude

taxeso

nim

movableprop

ertyor

netw

ealth

taxes

onthec

hangeo

fow

nershipof

prop

ertythroug

hinheritance

orgift

andtaxeso

nfin


ns

0283

5387

1463

106

Taxeso

nprop

ertyas

of


0844

14907

4013

274

SourceslowastEu

rosta

tlowastlowastWorld

Bank



DandEu

rosta

t

6 Complexity

minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3


Year2015201420132012

2011201020092008

200720062005

1200000

1400000

1600000

1800000

2000000

2200000

2400000

Fore

cast

pric

e


minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3


000

50000

100000

150000

200000

250000

300000

350000

400000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005




minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3


100000

120000

140000

160000

180000

200000

220000

240000

260000

Fore

cast

pric

eYear

2014201320122011

201020092008

200720062005





Complexity 7

minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

000

200000

400000

600000

800000

1000000

1200000

1400000

1600000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005


minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

150000

170000

190000

210000

230000

250000

270000

290000

310000

Fore

cast

pric

e

Year2015201420132012

201120102009

200820072006



minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

130000150000170000190000210000230000250000270000290000

Fore

cast

pric

e

Year2014201320122011

201020092008

200720062005


340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54


200000400000600000800000

100000012000001400000160000018000002000000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005


340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54


150000

200000

250000

300000

350000

400000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005


8 Complexity

340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54


Year201420132012

201120102009

200820072006

100000

120000

140000

160000

180000

200000

220000

240000

Fore

cast

pric

e


000

100000

200000

300000

400000

500000

600000

700000

800000

Fore

cast

pric

e

329

792

631

142

28

293

053

127

468

33

109

355

612

772

54

146

095

116

446

49

182

834

620

120

44

219

574

123

794

38

256

313

6

348

162

336

653

21

726

161

542

464

358

766

909

859

175

069

Net earningsYear

2015201420132012

20072006

2011201020092008

2005



4 Conclusion


Year2015201420132012

201120102009

200820072006

000

200000

400000

600000

800000

1000000

1200000

Fore

cast

pric

e

340

814

432

611

86

311

422

829

672

70

282

031

2

105

681

712

037

75

135

073

314

976

91

164

464

917

916

07

193

856

420

855

22

223

248

023

794

38

252

639

626

733

54

355

510

237

020

60

762

901

615

943

468

985

322

027

909

859

175

069

Net earnings


175

069

322

027

468

985

615

943

762

901

909

859

105

681

712

037

75

135

073

314

976

91

164

464

917

916

07

193

856

420

855

22

223

248

023

794

38

252

639

626

733

54

282

031

229

672

70

311

422

832

611

86

340

814

435

551

02

370

206

0

Net earningsYear

2014201320122011

201020092008

200720062005

000

500000

1000000

1500000

2000000

2500000

Fore

cast

pric

e




Complexity 9





References
































10 Complexity

























1 Introduction






2 Complexity

Envi

ronm

ent Tactical

management level


Strategicmanagement

level

Database Modelbase

Dialog

Users





2 Methodology





Complexity 3













4 Complexity

1

2

10



Wind speed




Training process


x1

x2

x3

x4

xm






o1(k+1) = t + Δt

o2(k+1) = t + Δt

o3(k+1) = t + Δt

o4(k+1) = t + Δt

om(k+1) = t + Δt

d1(k+1) = t + Δt

d2(k+1) = t + Δt

d3(k+1) = t + Δt

d4(k+1) = t + Δt

dm(k+1) = t + Δt






w1(k+1) w2(k+1) w3(k+1) w4(k+1) m(k+1) w











Complexity 5

(a) (b)

(c) (d)







Meteoalarm

Alert stages

Y R





6 Complexity




Decision support


Y R

Alert stages

Y R

Δt = 1

Δt = 1

Δt = 24

Δt = 1

Δt = 1 Δt = 24

Δt = 24

Δt = 24











Complexity 7



Windspeed



Airpressure

Airhumidity







Δ119905[h]





4 Conclusions


8 Complexity


10neurons

Data matrix




Input layer

Data matrix

Data matrix

Model response

Model response

ANN model



(156552 times 104 data)

Δt = 1 hour

Δt = 3 hours

Δt = 8 hours

Δt = 12 hours

Δt = 24 hours

Trainingprocess




Validationprocess

Evaluationprocess




1834 1836 1838 184 1842 1844 1846 1848 185Time (minutes)

2468

1012141618

Win

d sp

eed

(ms

)

Measurement

times105







Complexity 9




Acknowledgments


References



























10 Complexity
















1 Introduction







2 Complexity











Complexity 3













4 Complexity





Number ofprojects









Complexity 5















6 Complexity

















Complexity 7











neurons










neurons













8 Complexity



Type 1 12119908119879119908 + 119862119873sum119894=1

120585119894+ 119862 119873sum119894=1

120585lowast119894

119908119879120601(119909119894) + 119887 minus 119910

119894le 120576 + 120585lowast

119894119910119894minus 119908119879120601(119909

119894) minus 119887119894le 120576 + 120585

119894120585119894 120585lowast119894ge 0 119894 = 1119873

Type 2 12119908119879119908 minus 119862(]120576 + 1119873119873sum119894=1

(120585119894+ 120585lowast119894)) (119908119879120601(119909

119894) + 119887) minus 119910

119894le 120576 + 120585lowast

119894119910119894minus (119908119879120601(119909

119894) + 119887119894) le 120576 + 120585

119894120585119894 120585lowast119894ge 0 119894 = 1119873 120576 ge 0






SVM 1 20 0001 0083 2528 1547 SVM 2 20 0001 0091 2396 706 SVM 3 20 0001 0083 2921 2459SVM 4 20 0001 0091 3075 2277













Complexity 9




MAPE 4054 2697 2538 2688 1547 706




MAPE 3548 3022 2626 3516 2459 2277


5 Conclusions



10 Complexity

964

903

3537834

571

508

320

259

263

266341

1234


Amount of curbs




Preparation works

Earthworks

Drainage works


Other works


Project category

1071

697

3424

1209

628

538

366

238

241

280

1308


Amount of curbs




Preparation works

Earthworks

Drainage works


Other works

Project category


788713736

810842

698780

750971

916706

1290







739

854

600

787

1021

826

823

813

1156

792

1589


Amount of curbs




Preparation works

Earthworks

Drainage works


Other works

Project category






Complexity 11



Acknowledgments


References































12 Complexity

































Complexity 13






Editorial Board




Contents










and Andrej Soltesz5























2 Complexity










1 Introduction



















2 Complexity










3Complexity





















Z =wiZi =sumwiZi

sumwi= w1w1 +w2

Z1 +w2

w1 +w2Z1


1



N

M

Bel 3

el 2

el 1A

h = 3 m

b

d2

d2

y Tf = 20degC


4 Complexity








Inputlayer

Hiddenlayer 1

Hiddenlayer 2

Hiddenlayer 3

Hiddenlayer 4

Outputlayer

X

Y

Premiseparameters


A1

w1

w2

w1

w2

w2Z2

w1Z1

x1 x2

x1 x2

A2

B1

B2

N

N

120587

120587

f

f

sumZ


5Complexity











20

15

10Out

put

5

00 50 100 150 200



6 Complexity




10

8

Out

put

4

6

2

00 5 10 15 20

Data set index25 30


Testing data ()20

15

10

Out

put

5

00 10 20 30 40






7Complexity











20

15

Out

put

5

10

00 50 100 150 200 250

Index300 350


Testing data 20

15

Out

put

0

5

10

minus50 10 20 30 40 50 60

Index70 80

FIS output


8 Complexity




5 Conclusion





1580minus2006

1

2

3

4

5

6

7

8

30

Input[4040310503025]

Plot points101

Moveleft

Help Close

right down up


50 30 50 2 4 06 15 01 05 050


9Complexity


20

15

Out

put

5

10

00 5 10 15 20

Index25 30



Checking data

Columndimensions







3000 3000 200 100 010 02 588 553

3000 3000 200 100 030 03 214 187

3000 3000 300 100 010 05 294 344

3000 3000 300 100 040 04 132 132

3000 3000 400 100 040 01 208 183

4000 4000 200 100 010 02 776 798

4000 4000 200 100 040 0 382 383

4000 4000 300 100 020 04 356 374

4000 4000 300 100 030 05 216 228

4000 4000 300 100 040 0 38 377

4000 4000 400 100 010 01 1014 99

4000 4000 400 100 030 03 344 361

4000 4000 300 060 020 02 556 552

4000 4000 300 150 010 0 12 117

4000 4000 300 150 050 04 138 128

5000 5000 200 100 010 03 983 101

5000 5000 300 100 010 05 528 559

5000 5000 400 100 030 04 384 361

5000 5000 200 060 020 01 95 903

5000 5000 200 060 050 0 318 337

5000 5000 400 060 030 02 57 553

5000 5000 400 060 050 0 352 353

5000 5000 400 060 050 01 314 313

5000 5000 400 060 050 02 28 281

5000 5000 400 060 050 03 232 252

5000 5000 400 060 050 04 188 22

5000 5000 400 060 050 05 148 186

10 Complexity





Data Availability




References





















11Complexity


























12 Complexity









1 Introduction




















2 Methodology


2 Complexity









Envi

ronm

ent

Database Modelbase

Strategicmanagement

level



Dialog

Users

41

2 3


3Complexity












350 mm

Asphalt


Subgrade


4 Complexity






















2

Asphalt layer


3 Thickness (3 6 9 12 15 18 and 21 cm)


6







Input data x1

Input data xm

Inputlayer

Outputlayer

Hidden layers

Output data

⋮

⋮ ⋮


5Complexity













6 Complexity









A B C

Figure 5 Test cases

09940992


0965 0994

0998

0996

0997

0999


00

1

2

3

Pred

icte

d str

ess v

alue

(MPa

)

4

5

6

7



R2 = 0993

Case AEquality line


4 5 6 7


7Complexity





R2 = 09652

Pred

icte

d str

ess v

alue

s (M

Pa)

Case B

0

1

2

3

4

minus1

Equality line

0 1 2 3 4








(cm)


(MPa)


C2Modulus of



Stress 05 20ndash100 10 1400 and 9700




2



0

1

2

Stre

ss (M

Pa)

3

4


14 16 18 20


8 Complexity




4 Conclusions



Data Availability




References



0

1

2

Stre

ss (M

Pa)





0

1

2

Stre

ss (M

Pa)




05

15

25




Stre

ss (M

Pa)


9Complexity


























10 Complexity









1 Introduction











2 Complexity












119891 119883 997888rarr 119884 (1)



Complexity 3






Yes

No











surface







4 Complexity




(villages)



















Complexity 5


















119910 = 119891 (1199091 1199092 1199093 1199094 1199095 1199096 1199097) + 120576 (2)


119910 = 119891 (1199091 1199092 1199093 1199094 1199095 1199096 1199097) (3)


6 Complexity


119901 119910 1199091 1199092 1199093 1199094 1199095 1199096 11990975 5658 968 09 06 00 1968 6025 0013 1359 3292 09 05 6000 21420 00 0023 4276 1860 09 03 11160 780 375 1000037 4895 1860 01 03 2400 1000 00 0046 3230 800 09 05 1920 1818 1397 0059 1813 800 01 03 720 00 00 400067 19723 4131 09 05 3960 10960 2071 2586482 1611 1650 01 01 3000 00 00 0094 2500 1470 01 02 3445 939 385 00101 8005 1104 09 08 2959 14691 933 3746




1199091 133379 27500 560000 788491199092 059 010 090 0271199093 049 010 090 0161199094 30026 000 221200 345271199095 19316 000 214200 325841199096 10524 000 60250 121131199097 32429 000 300000 60326



119877 (119910 119910) = cov (119910 119910)120590푦120590푦


(4)









119867 asymp radic119873119872 (5)


NNP = NNW + NNB lt 119871 (6)


Complexity 7



119871 0901 0240 821 1455 1117119881 0879 0228 810 1088 578119879 0881 0233 805 1269 833

minus100

minus080

minus060

minus040

minus020

000

020

040

060

080

100


RT(y y)

RL(y y)

(a)

0

100

200

300

400

500

0 100 200 300 400 500

２－

３T

２－３L

(b)












8 Complexity

000

020

040

060

080

100

000 020 040 060 080 100

minus100

minus080

minus060

minus040


RV(y y)

RL(y y)

(a)

0

100

200

300

400

500

0 100 200 300 400 500

２－

３V

２－３L

(b)








SSE = sum푝isin퐿

(119910푝 minus 119910푝)2 (7)





Complexity 9


ANN 119877 (119910 119910) RMSE119871 119881 119879 119871 119881 119879MLPe-l 7-2-1






MLPs-l 7-5-1 119871 119881 119879MAPE


PEmax




100381610038161003816100381610038161003816100381610038161003816

PE푝 = 100381610038161003816100381610038161003816100381610038161003816119910푝 minus 119910푝119910푝

100381610038161003816100381610038161003816100381610038161003816 100PEmax = max (PE푝)

(8)



10 Complexity

0

500

1000

1500

2000

2500

3000

0 500 1000 1500 2000 2500 3000y




y













Acknowledgments


References








Complexity 11































1 Introduction





2 Complexity










Complexity 3











4 Complexity






119908(119897)new119894119895 = 119908(119897)old119894119895 minus 120578

120597120576119896

120597119908(119897)119894119895 (1)


119891 (119909) =1

1 + 119890minus119909 (2)


Δ119908(119897)119894119895 (119905) = minus120578120597120576119896

120597119908(119897)119894119895+ 120572Δ119908(119897)119894119895 (119905 minus 1) 0 lt 119905 lt 1 (3)





2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2005

Year

100000

300000

500000

700000

900000

1100000

Pric

e






Complexity 5

Table1Ba

sicinpu

tsinto

them

odel

theird

efinitio

nandbasic

characteris

tics

Varia

ble

Varia

bled

efinitio

n(according

tothes

ources


Minim

umMaxim

umMean

Standard

deviation

Unito

fmeasure

GDPlowast

GDP(grossdo

mestic

prod

uct)reflectsthe

totalvalue

ofallgoo

dsandservices

prod

uced

lessthe

valueo

fgoo

dsandservices

used


ptionin

theirp

rodu

ction

5142

3032820

47776356

7206493

5mileuro

BDPperc

apitalowast

GDPperc

apita

atmarketp

rices

iscalculated

asther

atio

ofGDPatmarketp

rices

tothea

verage

popu

latio

nof

aspecific

year

4200

84400

2382963

1575946

euro

RealGDPgrow

thratelowast

Thec

alculationof

thea

nnualgrowth

rateof


allowcomparis

onso

fthe

dynamicso

fecono

micdevelopm

entb

othover

timea

ndbetweenecon

omieso


119

163

384

Inequalityof

income

distr

ibutionlowast

Ther

atio

oftotalincom

ereceivedby

the2

0of

thep

opulationwith

theh

ighestincome(top

quintile)to

thatreceived

bythe2

0of

thep

opulationwith

thelow

estincom

e(lowestq

uintile)

32

83

483

116

-

Totalu


Unemploymentrates

representu

nemployed

person

sasa

percentage

ofthelabou

rforce

340

2750

907

432

Averagea

nnualn

etearningslowast

Netearnings

arec

alculated

from

grosse

arning

sbydedu

ctingthee


contrib

utions

andincometaxes

andadding

family

allowancesinthec

aseo

fhou

seho

ldsw

ithchild

ren

155077

3849018

1717581

1044515

euro

FDIlowastlowast


ityflo

wsinther

eportin

gecon

omyItis

thes

umof


ento

fearning

sandotherc

apita

lminus29679

734010

27948

67501

mil$

HICP-inflatio

nratelowast

Harmon

isedIndiceso

fCon

sumer

Prices

(HICPs)a

redesig

nedforinternatio

nalcom

paris

onso

fconsum

erpriceinfl

ation

minus16

153

238

220

VAT(

)lowastlowastlowast

Thev

alue


VATin

theE

urop

eanUnion

(EU)isa

generalbroadlybased

consum

ptiontaxassessed

onthev

alue

addedto

good

sand

services

1527

205

257

Taxeso

nprop

ertyas

of


Taxon

prop

ertyisdefin

edas

recurrentand

nonrecurrent

taxeso

ntheu

seownershipor

transfe

rof

prop

ertyTh

eseinclude

taxeso

nim

movableprop

ertyor

netw

ealth

taxes

onthec

hangeo

fow

nershipof

prop

ertythroug

hinheritance

orgift

andtaxeso

nfin


ns

0283

5387

1463

106

Taxeso

nprop

ertyas

of


0844

14907

4013

274

SourceslowastEu

rosta

tlowastlowastWorld

Bank



DandEu

rosta

t

6 Complexity

minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3


Year2015201420132012

2011201020092008

200720062005

1200000

1400000

1600000

1800000

2000000

2200000

2400000

Fore

cast

pric

e


minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3


000

50000

100000

150000

200000

250000

300000

350000

400000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005




minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3


100000

120000

140000

160000

180000

200000

220000

240000

260000

Fore

cast

pric

eYear

2014201320122011

201020092008

200720062005





Complexity 7

minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

000

200000

400000

600000

800000

1000000

1200000

1400000

1600000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005


minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

150000

170000

190000

210000

230000

250000

270000

290000

310000

Fore

cast

pric

e

Year2015201420132012

201120102009

200820072006



minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

130000150000170000190000210000230000250000270000290000

Fore

cast

pric

e

Year2014201320122011

201020092008

200720062005


340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54


200000400000600000800000

100000012000001400000160000018000002000000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005


340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54


150000

200000

250000

300000

350000

400000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005


8 Complexity

340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54


Year201420132012

201120102009

200820072006

100000

120000

140000

160000

180000

200000

220000

240000

Fore

cast

pric

e


000

100000

200000

300000

400000

500000

600000

700000

800000

Fore

cast

pric

e

329

792

631

142

28

293

053

127

468

33

109

355

612

772

54

146

095

116

446

49

182

834

620

120

44

219

574

123

794

38

256

313

6

348

162

336

653

21

726

161

542

464

358

766

909

859

175

069

Net earningsYear

2015201420132012

20072006

2011201020092008

2005



4 Conclusion


Year2015201420132012

201120102009

200820072006

000

200000

400000

600000

800000

1000000

1200000

Fore

cast

pric

e

340

814

432

611

86

311

422

829

672

70

282

031

2

105

681

712

037

75

135

073

314

976

91

164

464

917

916

07

193

856

420

855

22

223

248

023

794

38

252

639

626

733

54

355

510

237

020

60

762

901

615

943

468

985

322

027

909

859

175

069

Net earnings


175

069

322

027

468

985

615

943

762

901

909

859

105

681

712

037

75

135

073

314

976

91

164

464

917

916

07

193

856

420

855

22

223

248

023

794

38

252

639

626

733

54

282

031

229

672

70

311

422

832

611

86

340

814

435

551

02

370

206

0

Net earningsYear

2014201320122011

201020092008

200720062005

000

500000

1000000

1500000

2000000

2500000

Fore

cast

pric

e




Complexity 9





References
































10 Complexity

























1 Introduction






2 Complexity

Envi

ronm

ent Tactical

management level


Strategicmanagement

level

Database Modelbase

Dialog

Users





2 Methodology





Complexity 3













4 Complexity

1

2

10



Wind speed




Training process


x1

x2

x3

x4

xm






o1(k+1) = t + Δt

o2(k+1) = t + Δt

o3(k+1) = t + Δt

o4(k+1) = t + Δt

om(k+1) = t + Δt

d1(k+1) = t + Δt

d2(k+1) = t + Δt

d3(k+1) = t + Δt

d4(k+1) = t + Δt

dm(k+1) = t + Δt






w1(k+1) w2(k+1) w3(k+1) w4(k+1) m(k+1) w











Complexity 5

(a) (b)

(c) (d)







Meteoalarm

Alert stages

Y R





6 Complexity




Decision support


Y R

Alert stages

Y R

Δt = 1

Δt = 1

Δt = 24

Δt = 1

Δt = 1 Δt = 24

Δt = 24

Δt = 24











Complexity 7



Windspeed



Airpressure

Airhumidity







Δ119905[h]





4 Conclusions


8 Complexity


10neurons

Data matrix




Input layer

Data matrix

Data matrix

Model response

Model response

ANN model



(156552 times 104 data)

Δt = 1 hour

Δt = 3 hours

Δt = 8 hours

Δt = 12 hours

Δt = 24 hours

Trainingprocess




Validationprocess

Evaluationprocess




1834 1836 1838 184 1842 1844 1846 1848 185Time (minutes)

2468

1012141618

Win

d sp

eed

(ms

)

Measurement

times105







Complexity 9




Acknowledgments


References



























10 Complexity
















1 Introduction







2 Complexity











Complexity 3













4 Complexity





Number ofprojects









Complexity 5















6 Complexity

















Complexity 7











neurons










neurons













8 Complexity



Type 1 12119908119879119908 + 119862119873sum119894=1

120585119894+ 119862 119873sum119894=1

120585lowast119894

119908119879120601(119909119894) + 119887 minus 119910

119894le 120576 + 120585lowast

119894119910119894minus 119908119879120601(119909

119894) minus 119887119894le 120576 + 120585

119894120585119894 120585lowast119894ge 0 119894 = 1119873

Type 2 12119908119879119908 minus 119862(]120576 + 1119873119873sum119894=1

(120585119894+ 120585lowast119894)) (119908119879120601(119909

119894) + 119887) minus 119910

119894le 120576 + 120585lowast

119894119910119894minus (119908119879120601(119909

119894) + 119887119894) le 120576 + 120585

119894120585119894 120585lowast119894ge 0 119894 = 1119873 120576 ge 0






SVM 1 20 0001 0083 2528 1547 SVM 2 20 0001 0091 2396 706 SVM 3 20 0001 0083 2921 2459SVM 4 20 0001 0091 3075 2277













Complexity 9




MAPE 4054 2697 2538 2688 1547 706




MAPE 3548 3022 2626 3516 2459 2277


5 Conclusions



10 Complexity

964

903

3537834

571

508

320

259

263

266341

1234


Amount of curbs




Preparation works

Earthworks

Drainage works


Other works


Project category

1071

697

3424

1209

628

538

366

238

241

280

1308


Amount of curbs




Preparation works

Earthworks

Drainage works


Other works

Project category


788713736

810842

698780

750971

916706

1290







739

854

600

787

1021

826

823

813

1156

792

1589


Amount of curbs




Preparation works

Earthworks

Drainage works


Other works

Project category






Complexity 11



Acknowledgments


References































12 Complexity

































Complexity 13




Contents










and Andrej Soltesz5























2 Complexity










1 Introduction



















2 Complexity










3Complexity





















Z =wiZi =sumwiZi

sumwi= w1w1 +w2

Z1 +w2

w1 +w2Z1


1



N

M

Bel 3

el 2

el 1A

h = 3 m

b

d2

d2

y Tf = 20degC


4 Complexity








Inputlayer

Hiddenlayer 1

Hiddenlayer 2

Hiddenlayer 3

Hiddenlayer 4

Outputlayer

X

Y

Premiseparameters


A1

w1

w2

w1

w2

w2Z2

w1Z1

x1 x2

x1 x2

A2

B1

B2

N

N

120587

120587

f

f

sumZ


5Complexity











20

15

10Out

put

5

00 50 100 150 200



6 Complexity




10

8

Out

put

4

6

2

00 5 10 15 20

Data set index25 30


Testing data ()20

15

10

Out

put

5

00 10 20 30 40






7Complexity











20

15

Out

put

5

10

00 50 100 150 200 250

Index300 350


Testing data 20

15

Out

put

0

5

10

minus50 10 20 30 40 50 60

Index70 80

FIS output


8 Complexity




5 Conclusion





1580minus2006

1

2

3

4

5

6

7

8

30

Input[4040310503025]

Plot points101

Moveleft

Help Close

right down up


50 30 50 2 4 06 15 01 05 050


9Complexity


20

15

Out

put

5

10

00 5 10 15 20

Index25 30



Checking data

Columndimensions







3000 3000 200 100 010 02 588 553

3000 3000 200 100 030 03 214 187

3000 3000 300 100 010 05 294 344

3000 3000 300 100 040 04 132 132

3000 3000 400 100 040 01 208 183

4000 4000 200 100 010 02 776 798

4000 4000 200 100 040 0 382 383

4000 4000 300 100 020 04 356 374

4000 4000 300 100 030 05 216 228

4000 4000 300 100 040 0 38 377

4000 4000 400 100 010 01 1014 99

4000 4000 400 100 030 03 344 361

4000 4000 300 060 020 02 556 552

4000 4000 300 150 010 0 12 117

4000 4000 300 150 050 04 138 128

5000 5000 200 100 010 03 983 101

5000 5000 300 100 010 05 528 559

5000 5000 400 100 030 04 384 361

5000 5000 200 060 020 01 95 903

5000 5000 200 060 050 0 318 337

5000 5000 400 060 030 02 57 553

5000 5000 400 060 050 0 352 353

5000 5000 400 060 050 01 314 313

5000 5000 400 060 050 02 28 281

5000 5000 400 060 050 03 232 252

5000 5000 400 060 050 04 188 22

5000 5000 400 060 050 05 148 186

10 Complexity





Data Availability




References





















11Complexity


























12 Complexity









1 Introduction




















2 Methodology


2 Complexity









Envi

ronm

ent

Database Modelbase

Strategicmanagement

level



Dialog

Users

41

2 3


3Complexity












350 mm

Asphalt


Subgrade


4 Complexity






















2

Asphalt layer


3 Thickness (3 6 9 12 15 18 and 21 cm)


6







Input data x1

Input data xm

Inputlayer

Outputlayer

Hidden layers

Output data

⋮

⋮ ⋮


5Complexity













6 Complexity









A B C

Figure 5 Test cases

09940992


0965 0994

0998

0996

0997

0999


00

1

2

3

Pred

icte

d str

ess v

alue

(MPa

)

4

5

6

7



R2 = 0993

Case AEquality line


4 5 6 7


7Complexity





R2 = 09652

Pred

icte

d str

ess v

alue

s (M

Pa)

Case B

0

1

2

3

4

minus1

Equality line

0 1 2 3 4








(cm)


(MPa)


C2Modulus of



Stress 05 20ndash100 10 1400 and 9700




2



0

1

2

Stre

ss (M

Pa)

3

4


14 16 18 20


8 Complexity




4 Conclusions



Data Availability




References



0

1

2

Stre

ss (M

Pa)





0

1

2

Stre

ss (M

Pa)




05

15

25




Stre

ss (M

Pa)


9Complexity


























10 Complexity









1 Introduction











2 Complexity












119891 119883 997888rarr 119884 (1)



Complexity 3






Yes

No











surface







4 Complexity




(villages)



















Complexity 5


















119910 = 119891 (1199091 1199092 1199093 1199094 1199095 1199096 1199097) + 120576 (2)


119910 = 119891 (1199091 1199092 1199093 1199094 1199095 1199096 1199097) (3)


6 Complexity


119901 119910 1199091 1199092 1199093 1199094 1199095 1199096 11990975 5658 968 09 06 00 1968 6025 0013 1359 3292 09 05 6000 21420 00 0023 4276 1860 09 03 11160 780 375 1000037 4895 1860 01 03 2400 1000 00 0046 3230 800 09 05 1920 1818 1397 0059 1813 800 01 03 720 00 00 400067 19723 4131 09 05 3960 10960 2071 2586482 1611 1650 01 01 3000 00 00 0094 2500 1470 01 02 3445 939 385 00101 8005 1104 09 08 2959 14691 933 3746




1199091 133379 27500 560000 788491199092 059 010 090 0271199093 049 010 090 0161199094 30026 000 221200 345271199095 19316 000 214200 325841199096 10524 000 60250 121131199097 32429 000 300000 60326



119877 (119910 119910) = cov (119910 119910)120590푦120590푦


(4)









119867 asymp radic119873119872 (5)


NNP = NNW + NNB lt 119871 (6)


Complexity 7



119871 0901 0240 821 1455 1117119881 0879 0228 810 1088 578119879 0881 0233 805 1269 833

minus100

minus080

minus060

minus040

minus020

000

020

040

060

080

100


RT(y y)

RL(y y)

(a)

0

100

200

300

400

500

0 100 200 300 400 500

２－

３T

２－３L

(b)












8 Complexity

000

020

040

060

080

100

000 020 040 060 080 100

minus100

minus080

minus060

minus040


RV(y y)

RL(y y)

(a)

0

100

200

300

400

500

0 100 200 300 400 500

２－

３V

２－３L

(b)








SSE = sum푝isin퐿

(119910푝 minus 119910푝)2 (7)





Complexity 9


ANN 119877 (119910 119910) RMSE119871 119881 119879 119871 119881 119879MLPe-l 7-2-1






MLPs-l 7-5-1 119871 119881 119879MAPE


PEmax




100381610038161003816100381610038161003816100381610038161003816

PE푝 = 100381610038161003816100381610038161003816100381610038161003816119910푝 minus 119910푝119910푝

100381610038161003816100381610038161003816100381610038161003816 100PEmax = max (PE푝)

(8)



10 Complexity

0

500

1000

1500

2000

2500

3000

0 500 1000 1500 2000 2500 3000y




y













Acknowledgments


References








Complexity 11































1 Introduction





2 Complexity










Complexity 3











4 Complexity






119908(119897)new119894119895 = 119908(119897)old119894119895 minus 120578

120597120576119896

120597119908(119897)119894119895 (1)


119891 (119909) =1

1 + 119890minus119909 (2)


Δ119908(119897)119894119895 (119905) = minus120578120597120576119896

120597119908(119897)119894119895+ 120572Δ119908(119897)119894119895 (119905 minus 1) 0 lt 119905 lt 1 (3)





2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2005

Year

100000

300000

500000

700000

900000

1100000

Pric

e






Complexity 5

Table1Ba

sicinpu

tsinto

them

odel

theird

efinitio

nandbasic

characteris

tics

Varia

ble

Varia

bled

efinitio

n(according

tothes

ources


Minim

umMaxim

umMean

Standard

deviation

Unito

fmeasure

GDPlowast

GDP(grossdo

mestic

prod

uct)reflectsthe

totalvalue

ofallgoo

dsandservices

prod

uced

lessthe

valueo

fgoo

dsandservices

used


ptionin

theirp

rodu

ction

5142

3032820

47776356

7206493

5mileuro

BDPperc

apitalowast

GDPperc

apita

atmarketp

rices

iscalculated

asther

atio

ofGDPatmarketp

rices

tothea

verage

popu

latio

nof

aspecific

year

4200

84400

2382963

1575946

euro

RealGDPgrow

thratelowast

Thec

alculationof

thea

nnualgrowth

rateof


allowcomparis

onso

fthe

dynamicso

fecono

micdevelopm

entb

othover

timea

ndbetweenecon

omieso


119

163

384

Inequalityof

income

distr

ibutionlowast

Ther

atio

oftotalincom

ereceivedby

the2

0of

thep

opulationwith

theh

ighestincome(top

quintile)to

thatreceived

bythe2

0of

thep

opulationwith

thelow

estincom

e(lowestq

uintile)

32

83

483

116

-

Totalu


Unemploymentrates

representu

nemployed

person

sasa

percentage

ofthelabou

rforce

340

2750

907

432

Averagea

nnualn

etearningslowast

Netearnings

arec

alculated

from

grosse

arning

sbydedu

ctingthee


contrib

utions

andincometaxes

andadding

family

allowancesinthec

aseo

fhou

seho

ldsw

ithchild

ren

155077

3849018

1717581

1044515

euro

FDIlowastlowast


ityflo

wsinther

eportin

gecon

omyItis

thes

umof


ento

fearning

sandotherc

apita

lminus29679

734010

27948

67501

mil$

HICP-inflatio

nratelowast

Harmon

isedIndiceso

fCon

sumer

Prices

(HICPs)a

redesig

nedforinternatio

nalcom

paris

onso

fconsum

erpriceinfl

ation

minus16

153

238

220

VAT(

)lowastlowastlowast

Thev

alue


VATin

theE

urop

eanUnion

(EU)isa

generalbroadlybased

consum

ptiontaxassessed

onthev

alue

addedto

good

sand

services

1527

205

257

Taxeso

nprop

ertyas

of


Taxon

prop

ertyisdefin

edas

recurrentand

nonrecurrent

taxeso

ntheu

seownershipor

transfe

rof

prop

ertyTh

eseinclude

taxeso

nim

movableprop

ertyor

netw

ealth

taxes

onthec

hangeo

fow

nershipof

prop

ertythroug

hinheritance

orgift

andtaxeso

nfin


ns

0283

5387

1463

106

Taxeso

nprop

ertyas

of


0844

14907

4013

274

SourceslowastEu

rosta

tlowastlowastWorld

Bank



DandEu

rosta

t

6 Complexity

minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3


Year2015201420132012

2011201020092008

200720062005

1200000

1400000

1600000

1800000

2000000

2200000

2400000

Fore

cast

pric

e


minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3


000

50000

100000

150000

200000

250000

300000

350000

400000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005




minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3


100000

120000

140000

160000

180000

200000

220000

240000

260000

Fore

cast

pric

eYear

2014201320122011

201020092008

200720062005





Complexity 7

minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

000

200000

400000

600000

800000

1000000

1200000

1400000

1600000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005


minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

150000

170000

190000

210000

230000

250000

270000

290000

310000

Fore

cast

pric

e

Year2015201420132012

201120102009

200820072006



minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

130000150000170000190000210000230000250000270000290000

Fore

cast

pric

e

Year2014201320122011

201020092008

200720062005


340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54


200000400000600000800000

100000012000001400000160000018000002000000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005


340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54


150000

200000

250000

300000

350000

400000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005


8 Complexity

340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54


Year201420132012

201120102009

200820072006

100000

120000

140000

160000

180000

200000

220000

240000

Fore

cast

pric

e


000

100000

200000

300000

400000

500000

600000

700000

800000

Fore

cast

pric

e

329

792

631

142

28

293

053

127

468

33

109

355

612

772

54

146

095

116

446

49

182

834

620

120

44

219

574

123

794

38

256

313

6

348

162

336

653

21

726

161

542

464

358

766

909

859

175

069

Net earningsYear

2015201420132012

20072006

2011201020092008

2005



4 Conclusion


Year2015201420132012

201120102009

200820072006

000

200000

400000

600000

800000

1000000

1200000

Fore

cast

pric

e

340

814

432

611

86

311

422

829

672

70

282

031

2

105

681

712

037

75

135

073

314

976

91

164

464

917

916

07

193

856

420

855

22

223

248

023

794

38

252

639

626

733

54

355

510

237

020

60

762

901

615

943

468

985

322

027

909

859

175

069

Net earnings


175

069

322

027

468

985

615

943

762

901

909

859

105

681

712

037

75

135

073

314

976

91

164

464

917

916

07

193

856

420

855

22

223

248

023

794

38

252

639

626

733

54

282

031

229

672

70

311

422

832

611

86

340

814

435

551

02

370

206

0

Net earningsYear

2014201320122011

201020092008

200720062005

000

500000

1000000

1500000

2000000

2500000

Fore

cast

pric

e




Complexity 9





References
































10 Complexity

























1 Introduction






2 Complexity

Envi

ronm

ent Tactical

management level


Strategicmanagement

level

Database Modelbase

Dialog

Users





2 Methodology





Complexity 3













4 Complexity

1

2

10



Wind speed




Training process


x1

x2

x3

x4

xm






o1(k+1) = t + Δt

o2(k+1) = t + Δt

o3(k+1) = t + Δt

o4(k+1) = t + Δt

om(k+1) = t + Δt

d1(k+1) = t + Δt

d2(k+1) = t + Δt

d3(k+1) = t + Δt

d4(k+1) = t + Δt

dm(k+1) = t + Δt






w1(k+1) w2(k+1) w3(k+1) w4(k+1) m(k+1) w











Complexity 5

(a) (b)

(c) (d)







Meteoalarm

Alert stages

Y R





6 Complexity




Decision support


Y R

Alert stages

Y R

Δt = 1

Δt = 1

Δt = 24

Δt = 1

Δt = 1 Δt = 24

Δt = 24

Δt = 24











Complexity 7



Windspeed



Airpressure

Airhumidity







Δ119905[h]





4 Conclusions


8 Complexity


10neurons

Data matrix




Input layer

Data matrix

Data matrix

Model response

Model response

ANN model



(156552 times 104 data)

Δt = 1 hour

Δt = 3 hours

Δt = 8 hours

Δt = 12 hours

Δt = 24 hours

Trainingprocess




Validationprocess

Evaluationprocess




1834 1836 1838 184 1842 1844 1846 1848 185Time (minutes)

2468

1012141618

Win

d sp

eed

(ms

)

Measurement

times105







Complexity 9




Acknowledgments


References



























10 Complexity
















1 Introduction







2 Complexity











Complexity 3













4 Complexity





Number ofprojects









Complexity 5















6 Complexity

















Complexity 7











neurons










neurons













8 Complexity



Type 1 12119908119879119908 + 119862119873sum119894=1

120585119894+ 119862 119873sum119894=1

120585lowast119894

119908119879120601(119909119894) + 119887 minus 119910

119894le 120576 + 120585lowast

119894119910119894minus 119908119879120601(119909

119894) minus 119887119894le 120576 + 120585

119894120585119894 120585lowast119894ge 0 119894 = 1119873

Type 2 12119908119879119908 minus 119862(]120576 + 1119873119873sum119894=1

(120585119894+ 120585lowast119894)) (119908119879120601(119909

119894) + 119887) minus 119910

119894le 120576 + 120585lowast

119894119910119894minus (119908119879120601(119909

119894) + 119887119894) le 120576 + 120585

119894120585119894 120585lowast119894ge 0 119894 = 1119873 120576 ge 0






SVM 1 20 0001 0083 2528 1547 SVM 2 20 0001 0091 2396 706 SVM 3 20 0001 0083 2921 2459SVM 4 20 0001 0091 3075 2277













Complexity 9




MAPE 4054 2697 2538 2688 1547 706




MAPE 3548 3022 2626 3516 2459 2277


5 Conclusions



10 Complexity

964

903

3537834

571

508

320

259

263

266341

1234


Amount of curbs




Preparation works

Earthworks

Drainage works


Other works


Project category

1071

697

3424

1209

628

538

366

238

241

280

1308


Amount of curbs




Preparation works

Earthworks

Drainage works


Other works

Project category


788713736

810842

698780

750971

916706

1290







739

854

600

787

1021

826

823

813

1156

792

1589


Amount of curbs




Preparation works

Earthworks

Drainage works


Other works

Project category






Complexity 11



Acknowledgments


References































12 Complexity

































Complexity 13






and Andrej Soltesz5























2 Complexity










1 Introduction



















2 Complexity










3Complexity





















Z =wiZi =sumwiZi

sumwi= w1w1 +w2

Z1 +w2

w1 +w2Z1


1



N

M

Bel 3

el 2

el 1A

h = 3 m

b

d2

d2

y Tf = 20degC


4 Complexity








Inputlayer

Hiddenlayer 1

Hiddenlayer 2

Hiddenlayer 3

Hiddenlayer 4

Outputlayer

X

Y

Premiseparameters


A1

w1

w2

w1

w2

w2Z2

w1Z1

x1 x2

x1 x2

A2

B1

B2

N

N

120587

120587

f

f

sumZ


5Complexity











20

15

10Out

put

5

00 50 100 150 200



6 Complexity




10

8

Out

put

4

6

2

00 5 10 15 20

Data set index25 30


Testing data ()20

15

10

Out

put

5

00 10 20 30 40






7Complexity











20

15

Out

put

5

10

00 50 100 150 200 250

Index300 350


Testing data 20

15

Out

put

0

5

10

minus50 10 20 30 40 50 60

Index70 80

FIS output


8 Complexity




5 Conclusion





1580minus2006

1

2

3

4

5

6

7

8

30

Input[4040310503025]

Plot points101

Moveleft

Help Close

right down up


50 30 50 2 4 06 15 01 05 050


9Complexity


20

15

Out

put

5

10

00 5 10 15 20

Index25 30



Checking data

Columndimensions







3000 3000 200 100 010 02 588 553

3000 3000 200 100 030 03 214 187

3000 3000 300 100 010 05 294 344

3000 3000 300 100 040 04 132 132

3000 3000 400 100 040 01 208 183

4000 4000 200 100 010 02 776 798

4000 4000 200 100 040 0 382 383

4000 4000 300 100 020 04 356 374

4000 4000 300 100 030 05 216 228

4000 4000 300 100 040 0 38 377

4000 4000 400 100 010 01 1014 99

4000 4000 400 100 030 03 344 361

4000 4000 300 060 020 02 556 552

4000 4000 300 150 010 0 12 117

4000 4000 300 150 050 04 138 128

5000 5000 200 100 010 03 983 101

5000 5000 300 100 010 05 528 559

5000 5000 400 100 030 04 384 361

5000 5000 200 060 020 01 95 903

5000 5000 200 060 050 0 318 337

5000 5000 400 060 030 02 57 553

5000 5000 400 060 050 0 352 353

5000 5000 400 060 050 01 314 313

5000 5000 400 060 050 02 28 281

5000 5000 400 060 050 03 232 252

5000 5000 400 060 050 04 188 22

5000 5000 400 060 050 05 148 186

10 Complexity





Data Availability




References





















11Complexity


























12 Complexity









1 Introduction




















2 Methodology


2 Complexity









Envi

ronm

ent

Database Modelbase

Strategicmanagement

level



Dialog

Users

41

2 3


3Complexity












350 mm

Asphalt


Subgrade


4 Complexity






















2

Asphalt layer


3 Thickness (3 6 9 12 15 18 and 21 cm)


6







Input data x1

Input data xm

Inputlayer

Outputlayer

Hidden layers

Output data

⋮

⋮ ⋮


5Complexity













6 Complexity









A B C

Figure 5 Test cases

09940992


0965 0994

0998

0996

0997

0999


00

1

2

3

Pred

icte

d str

ess v

alue

(MPa

)

4

5

6

7



R2 = 0993

Case AEquality line


4 5 6 7


7Complexity





R2 = 09652

Pred

icte

d str

ess v

alue

s (M

Pa)

Case B

0

1

2

3

4

minus1

Equality line

0 1 2 3 4








(cm)


(MPa)


C2Modulus of



Stress 05 20ndash100 10 1400 and 9700




2



0

1

2

Stre

ss (M

Pa)

3

4


14 16 18 20


8 Complexity




4 Conclusions



Data Availability




References



0

1

2

Stre

ss (M

Pa)





0

1

2

Stre

ss (M

Pa)




05

15

25




Stre

ss (M

Pa)


9Complexity


























10 Complexity









1 Introduction











2 Complexity












119891 119883 997888rarr 119884 (1)



Complexity 3






Yes

No











surface







4 Complexity




(villages)



















Complexity 5


















119910 = 119891 (1199091 1199092 1199093 1199094 1199095 1199096 1199097) + 120576 (2)


119910 = 119891 (1199091 1199092 1199093 1199094 1199095 1199096 1199097) (3)


6 Complexity


119901 119910 1199091 1199092 1199093 1199094 1199095 1199096 11990975 5658 968 09 06 00 1968 6025 0013 1359 3292 09 05 6000 21420 00 0023 4276 1860 09 03 11160 780 375 1000037 4895 1860 01 03 2400 1000 00 0046 3230 800 09 05 1920 1818 1397 0059 1813 800 01 03 720 00 00 400067 19723 4131 09 05 3960 10960 2071 2586482 1611 1650 01 01 3000 00 00 0094 2500 1470 01 02 3445 939 385 00101 8005 1104 09 08 2959 14691 933 3746




1199091 133379 27500 560000 788491199092 059 010 090 0271199093 049 010 090 0161199094 30026 000 221200 345271199095 19316 000 214200 325841199096 10524 000 60250 121131199097 32429 000 300000 60326



119877 (119910 119910) = cov (119910 119910)120590푦120590푦


(4)









119867 asymp radic119873119872 (5)


NNP = NNW + NNB lt 119871 (6)


Complexity 7



119871 0901 0240 821 1455 1117119881 0879 0228 810 1088 578119879 0881 0233 805 1269 833

minus100

minus080

minus060

minus040

minus020

000

020

040

060

080

100


RT(y y)

RL(y y)

(a)

0

100

200

300

400

500

0 100 200 300 400 500

２－

３T

２－３L

(b)












8 Complexity

000

020

040

060

080

100

000 020 040 060 080 100

minus100

minus080

minus060

minus040


RV(y y)

RL(y y)

(a)

0

100

200

300

400

500

0 100 200 300 400 500

２－

３V

２－３L

(b)








SSE = sum푝isin퐿

(119910푝 minus 119910푝)2 (7)





Complexity 9


ANN 119877 (119910 119910) RMSE119871 119881 119879 119871 119881 119879MLPe-l 7-2-1






MLPs-l 7-5-1 119871 119881 119879MAPE


PEmax




100381610038161003816100381610038161003816100381610038161003816

PE푝 = 100381610038161003816100381610038161003816100381610038161003816119910푝 minus 119910푝119910푝

100381610038161003816100381610038161003816100381610038161003816 100PEmax = max (PE푝)

(8)



10 Complexity

0

500

1000

1500

2000

2500

3000

0 500 1000 1500 2000 2500 3000y




y













Acknowledgments


References








Complexity 11































1 Introduction





2 Complexity










Complexity 3











4 Complexity






119908(119897)new119894119895 = 119908(119897)old119894119895 minus 120578

120597120576119896

120597119908(119897)119894119895 (1)


119891 (119909) =1

1 + 119890minus119909 (2)


Δ119908(119897)119894119895 (119905) = minus120578120597120576119896

120597119908(119897)119894119895+ 120572Δ119908(119897)119894119895 (119905 minus 1) 0 lt 119905 lt 1 (3)





2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2005

Year

100000

300000

500000

700000

900000

1100000

Pric

e






Complexity 5

Table1Ba

sicinpu

tsinto

them

odel

theird

efinitio

nandbasic

characteris

tics

Varia

ble

Varia

bled

efinitio

n(according

tothes

ources


Minim

umMaxim

umMean

Standard

deviation

Unito

fmeasure

GDPlowast

GDP(grossdo

mestic

prod

uct)reflectsthe

totalvalue

ofallgoo

dsandservices

prod

uced

lessthe

valueo

fgoo

dsandservices

used


ptionin

theirp

rodu

ction

5142

3032820

47776356

7206493

5mileuro

BDPperc

apitalowast

GDPperc

apita

atmarketp

rices

iscalculated

asther

atio

ofGDPatmarketp

rices

tothea

verage

popu

latio

nof

aspecific

year

4200

84400

2382963

1575946

euro

RealGDPgrow

thratelowast

Thec

alculationof

thea

nnualgrowth

rateof


allowcomparis

onso

fthe

dynamicso

fecono

micdevelopm

entb

othover

timea

ndbetweenecon

omieso


119

163

384

Inequalityof

income

distr

ibutionlowast

Ther

atio

oftotalincom

ereceivedby

the2

0of

thep

opulationwith

theh

ighestincome(top

quintile)to

thatreceived

bythe2

0of

thep

opulationwith

thelow

estincom

e(lowestq

uintile)

32

83

483

116

-

Totalu


Unemploymentrates

representu

nemployed

person

sasa

percentage

ofthelabou

rforce

340

2750

907

432

Averagea

nnualn

etearningslowast

Netearnings

arec

alculated

from

grosse

arning

sbydedu

ctingthee


contrib

utions

andincometaxes

andadding

family

allowancesinthec

aseo

fhou

seho

ldsw

ithchild

ren

155077

3849018

1717581

1044515

euro

FDIlowastlowast


ityflo

wsinther

eportin

gecon

omyItis

thes

umof


ento

fearning

sandotherc

apita

lminus29679

734010

27948

67501

mil$

HICP-inflatio

nratelowast

Harmon

isedIndiceso

fCon

sumer

Prices

(HICPs)a

redesig

nedforinternatio

nalcom

paris

onso

fconsum

erpriceinfl

ation

minus16

153

238

220

VAT(

)lowastlowastlowast

Thev

alue


VATin

theE

urop

eanUnion

(EU)isa

generalbroadlybased

consum

ptiontaxassessed

onthev

alue

addedto

good

sand

services

1527

205

257

Taxeso

nprop

ertyas

of


Taxon

prop

ertyisdefin

edas

recurrentand

nonrecurrent

taxeso

ntheu

seownershipor

transfe

rof

prop

ertyTh

eseinclude

taxeso

nim

movableprop

ertyor

netw

ealth

taxes

onthec

hangeo

fow

nershipof

prop

ertythroug

hinheritance

orgift

andtaxeso

nfin


ns

0283

5387

1463

106

Taxeso

nprop

ertyas

of


0844

14907

4013

274

SourceslowastEu

rosta

tlowastlowastWorld

Bank



DandEu

rosta

t

6 Complexity

minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3


Year2015201420132012

2011201020092008

200720062005

1200000

1400000

1600000

1800000

2000000

2200000

2400000

Fore

cast

pric

e


minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3


000

50000

100000

150000

200000

250000

300000

350000

400000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005




minus14

80minus1

373

minus12

66minus1

160

minus10

53minus9

46

minus83

9minus7

32

minus62

6minus5

19

minus41

2minus3

05

minus19

8minus0

92

015

122

229

336

442

549

656

763

870

976

108

3


100000

120000

140000

160000

180000

200000

220000

240000

260000

Fore

cast

pric

eYear

2014201320122011

201020092008

200720062005





Complexity 7

minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

000

200000

400000

600000

800000

1000000

1200000

1400000

1600000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005


minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

150000

170000

190000

210000

230000

250000

270000

290000

310000

Fore

cast

pric

e

Year2015201420132012

201120102009

200820072006



minus16

0minus0

92

minus02

50

431

101

782

463

133

814

485

165

846

517

197

868

549

229

8910

57

112

411

92

126

013

27

139

514

62

HICP-inflation rate

130000150000170000190000210000230000250000270000290000

Fore

cast

pric

e

Year2014201320122011

201020092008

200720062005


340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54


200000400000600000800000

100000012000001400000160000018000002000000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005


340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54


150000

200000

250000

300000

350000

400000

Fore

cast

pric

e

Year2015201420132012

2011201020092008

200720062005


8 Complexity

340

436

533

629

726

822

918

101

511

11

120

813

04

140

014

97

159

316

90

178

618

82

197

920

75

217

222

68

236

424

61

255

726

54


Year201420132012

201120102009

200820072006

100000

120000

140000

160000

180000

200000

220000

240000

Fore

cast

pric

e


000

100000

200000

300000

400000

500000

600000

700000

800000

Fore

cast

pric

e

329

792

631

142

28

293

053

127

468

33

109

355

612

772

54

146

095

116

446

49

182

834

620

120

44

219

574

123

794

38

256

313

6

348

162

336

653

21

726

161

542

464

358

766

909

859

175

069

Net earningsYear

2015201420132012

20072006

2011201020092008

2005



4 Conclusion


Year2015201420132012

201120102009

200820072006

000

200000

400000

600000

800000

1000000

1200000

Fore

cast

pric

e

340

814

432

611

86

311

422

829

672

70

282

031

2

105

681

712

037

75

135

073

314

976

91

164

464

917

916

07

193

856

420

855

22

223

248

023

794

38

252

639

626

733

54

355

510

237

020

60

762

901

615

943

468

985

322

027

909

859

175

069

Net earnings


175

069

322

027

468

985

615

943

762

901

909

859

105

681

712

037

75

135

073

314

976

91

164

464

917

916

07

193

856

420

855

22

223

248

023

794

38

252

639

626

733

54

282

031

229

672

70

311

422

832

611

86

340

814

435

551

02

370

206

0

Net earningsYear

2014201320122011

201020092008

200720062005

000

500000

1000000

1500000

2000000

2500000

Fore

cast

pric

e




Complexity 9





References
































10 Complexity

























1 Introduction






2 Complexity

Envi

ronm

ent Tactical

management level


Strategicmanagement

level

Database Modelbase

Dialog

Users





2 Methodology





Complexity 3













4 Complexity

1

2

10



Wind speed




Training process


x1

x2

x3

x4

xm






o1(k+1) = t + Δt

o2(k+1) = t + Δt

o3(k+1) = t + Δt

o4(k+1) = t + Δt

om(k+1) = t + Δt

d1(k+1) = t + Δt

d2(k+1) = t + Δt

d3(k+1) = t + Δt

d4(k+1) = t + Δt

dm(k+1) = t + Δt






w1(k+1) w2(k+1) w3(k+1) w4(k+1) m(k+1) w











Complexity 5

(a) (b)

(c) (d)







Meteoalarm

Alert stages

Y R





6 Complexity




Decision support


Y R

Alert stages

Y R

Δt = 1

Δt = 1

Δt = 24

Δt = 1

Δt = 1 Δt = 24

Δt = 24

Δt = 24











Complexity 7



Windspeed



Airpressure

Airhumidity







Δ119905[h]





4 Conclusions


8 Complexity


10neurons

Data matrix




Input layer

Data matrix

Data matrix

Model response

Model response

ANN model



(156552 times 104 data)

Δt = 1 hour

Δt = 3 hours

Δt = 8 hours

Δt = 12 hours

Δt = 24 hours

Trainingprocess




Validationprocess

Evaluationprocess




1834 1836 1838 184 1842 1844 1846 1848 185Time (minutes)

2468

1012141618

Win

d sp

eed

(ms

)

Measurement

times105







Complexity 9




Acknowledgments


References



























10 Complexity
















1 Introduction







2 Complexity











Complexity 3













4 Complexity





Number ofprojects









Complexity 5















6 Complexity

















Complexity 7











neurons










neurons













8 Complexity



Type 1 12119908119879119908 + 119862119873sum119894=1

120585119894+ 119862 119873sum119894=1

120585lowast119894

119908119879120601(119909119894) + 119887 minus 119910

119894le 120576 + 120585lowast

119894119910119894minus 119908119879120601(119909

119894) minus 119887119894le 120576 + 120585

119894120585119894 120585lowast119894ge 0 119894 = 1119873

Type 2 12119908119879119908 minus 119862(]120576 + 1119873119873sum119894=1

(120585119894+ 120585lowast119894)) (119908119879120601(119909

119894) + 119887) minus 119910

119894le 120576 + 120585lowast

119894119910119894minus (119908119879120601(119909

119894) + 119887119894) le 120576 + 120585

119894120585119894 120585lowast119894ge 0 119894 = 1119873 120576 ge 0






SVM 1 20 0001 0083 2528 1547 SVM 2 20 0001 0091 2396 706 SVM 3 20 0001 0083 2921 2459SVM 4 20 0001 0091 3075 2277













Complexity 9




MAPE 4054 2697 2538 2688 1547 706




MAPE 3548 3022 2626 3516 2459 2277


5 Conclusions



10 Complexity

964

903

3537834

571

508

320

259

263

266341

1234


Amount of curbs




Preparation works

Earthworks

Drainage works


Other works


Project category

1071

697

3424

1209

628

538

366

238

241

280

1308


Amount of curbs




Preparation works

Earthworks

Drainage works


Other works

Project category


788713736

810842

698780

750971

916706

1290







739

854

600

787

1021

826

823

813

1156

792

1589


Amount of curbs




Preparation works

Earthworks

Drainage works


Other works

Project category






Complexity 11



Acknowledgments


References































12 Complexity

































Complexity 13




Artiﬁcial Neural Networks and Fuzzy Neural Networks for ...

Documents

Transcript of Artiﬁcial Neural Networks and Fuzzy Neural Networks for ...