Abstract: — Software project management begins with a set ofcollected activities called project planning activity. Beforestarting a project, the program’s team must evaluate the work tobe done, the resources to be reorganized, and the time from thebeginning computation. When these activities have beencompleted, the program’s team should establish a set of projectsthat will assign program engineering tasks, key milestones,determine the responsibly for each task, and identify associateddependencies among participants that may have a strong impacton progress. In general, there isn’t a complete accurate estimationmethod, but in this research I tried to discover the bestprogramming methods to find the best programming estimate.The aim of this research is to present a study of the principles forreducing the cost of software and understanding how thesetechniques are applied to general program divisions. We providebasic algorithms in Artificial Intelligence, Artificial NeuralNetworks, Genetic Algorithms, and Fuzzy Logic Algorithm todetermine which algorithm is the most appropriate to find thebest estimates as possible in terms of precision. The best resultswere found in Neural Networks, but competitive results werefound between types of Neural Networks (FFNN, CNN, ENN,RBFN and NARX). The NARX network was observed to providethe best accuracy, but Genetic Algorithm proved better thanFuzzy Logic which is the worst compared to Neural Networksand Genetic Algorithms.Keywords: —Network Architecture, Neural Networks, FuzzyNeural Network, Genetic Algorithms, Fuzzy Logic.I. INTRODUCTION1The efficiency of the software project assessment hasbecome one of the most important challenges and the mostchallenging activities for software developers, it is alsodifficult to achieve accuracy during software development.Project planning and control of development principles are notpossible unless the reliability of the estimate and the ability tofocus on the calculation of the effort are met. The timerequired to complete the project and the cost of the sources,including the size of employees, while taking into account thedetermination of the estimation levels, whether over estimateor under estimate, and the accuracy of size of for the softwarebeing built, all depends first on the extent of efficiencyestimation sources staff time and cost, based on the principleof artificial intelligent neural networks, Genetic Algorithmsand Fuzzy logic. The first researcher Esra’a Z.M.Quba[7]made a comparison between the standard model of estimationCOCOMO and four models of Artificial Neural Networks, thesecond researcher, Taghreed Riyadh[10], conducted acomparative study between what the first researcher is tacklingwith the Neural Networks and the COCOMO model comparedto the genetic algorithm, the third researcher, MuhammadYahya[6] presented the Fuzzy Logic in estimation.Previous work:The researchers used Tsakonas and Dounias in 2009Genetic programming to give high-accuracy mathematicianand to help find the estimator’s effort to give a relationshipbetween properties project and work required data, which usedCOCOMO NASA dataset and COC81 dataset, the results werecompared where they showed excellent geneticprogramming[11]. In 2013 the two researchers SultanAljahdali, Alaa F.Sheta developed two models to estimatesoftware effort using Multigene Symbolic Regression GeneticProgramming, the models have been used in one (SLOC) as aninput variable to calculate effort and system Inputs, outputs,files and User Inquiries for estimation FP In the second model,the model has proved its efficiency compared to other models[5].1.1 Estimation the programming effort and methods:The process of estimating a project is not a static science,but it is a combination of historical data with data extracted bytechniques to demonstrate the accuracy of matching estimationto achieve a precise guess that requires some important thingsto accomplish the objectives of the project successfully.Although the estimation alone is of great importance insoftware, a plan must be developed for the project andFirdews A.Abulalqader * Aseel W.AliM.S.c. student Assistant pro.University of Al Mosul University of Al MosulSoftware Engineering Department Software Engineering DepartmentMosul, Iraq Mosul, IraqE-mail: Firdewsalsalman@yahoo.com E-mail: AseelW.Ali71@yahoo.comComparing Different Estimation Methods forSoftware EffortTU”OOVBM*OUFSOBUJPOBM$POGFSFODFPO*OGPSNBUJPOBOE4DJFODFT “*$*4¥*&&&%0*”J$*4Authorized licensed use limited to: Federation University Australia. Downloaded on May 31,2021 at 01:10:53 UTC from IEEE Xplore. Restrictions apply.knowledge of the risks that may be faced by the developerduring the process of development and construction and theextent of control, which in some projects aspires developers toestimate the process at early stages of the life cycle of theproject. Thus, the estimation is insufficient, for example (thedesigner cannot estimate the effort, scheduling, costeffectively), so, this will lead to the confusion of the work ofthe project manager and may not achieve reliability in theestimation. For these reasons, there are some cases advised tobe followed:-x Estimation depends on similar and completed projects(realization of completed projects). This option maybe reasonable if the current project is similar to acomplete company (customers, working conditions,environment, and deadline).x Delay the estimation to late stages of the project togather as much information as possible and know theproject requirements and specifications. This optionis often impractical and unrealistic and the costestimate must be made in advance.x Use simple splitting techniques when pursuing theestimation process.x Use one or more experimental models in the estimationprocess.x For the last two options, they are effective methods ofestimation. Divide-and-Conquer techniques dividethe project into basic functions and related events.The process of estimation time, effort,and cost.2.1 The collection and classification of the database:For the first researcher, the programming effort estimated isbased on a historical database ready for 60 projects of theNASA space agency, where each project contains 15 valuesrepresenting the cost factors. This data has executed,completed and calculated the program effort for each projectin its accurate actual effort estimation. Figure (1) shows theNASA database diagram, and table (1) shows NASA datacollection used in research with numerical values.01000200030004000Effort in (PM)Project No.1 4 10 12 18 22 28 34 40 44 48 50 53 58 60Figure (1): The NASA database diagram.2.1.1 Use the COCOMO standard in calculationprogrammatic estimation:The COCOMO model is Empirical, which was developedby compiling vast data from a large number of softwareprojects. This data was analyzed to discover the equations thatwere most appropriate for the observations reaches, theseequations relate the size of the system and characteristics ofthe project, product and development team, with the effortrequired to develop the system, the COCOMO model istherefore used by a large number of project managers. Unlikeother cost estimation models, COCOMO is an open model andtherefore all its details are published, based on the COCOMOprinciple, the following results were obtained, Table(2) showsresults of the COCOMO method in estimating theprogramming effort.Table (1): NASA data collection used in research withnumerical values.Table (2): COCOMO results.Table (3): measurements of COCOMO results. Estimator Effort using the COCOMOmediumMeasureMMRE2.5517RMSE 136.6750BRE 0.3847VAF 88.9870 Project… 60Project…. 35Project1 RELY 1.15 ….. 1.15 ….. 1.15DATA 1.00 ….. 1.00 ….. 0.94CPLX 1.15 ….. 1.15 ….. 1.15TIME 1.11 ….. 1.00 ….. 1.00STOR 1.06 ….. 1.00 ….. 1.00VIRT 0.87 ….. 1.00 ….. 0.87TURN 1.07 ….. 1.00 ….. 0.87ACAP 1.00 ….. 1.00 ….. 1.00AEXP 0.91 ….. 0.91 ….. 1.00PCAP 1.00 ….. 0.86 ….. 1.00VEXP 1.00 ….. 1.00 ….. 1.00LEXP 1.00 ….. 1.00 ….. 0.95MODP 1.10 ….. 1.00 ….. 0.91TOOL 0.83 ….. 1.00 ….. 1.00SCED 1.00 ….. 1.00 ….. 1.08KSLOC 370 ….. 38 ….. 2.2Actual effort 3240 ….. 210 ….. 8.4(months) NO ProjectEffort EstimatorEffort Real51 57.8513 6252 379.3830 30053 34.0296 4854 11.0607 10.855 142.7862 12056 234.1576 37057 91.9847 6058 182.5637 21059 892.6617 124860 42.1172 72 NO ProjectCost factorsAuthorized licensed use limited to: Federation University Australia. Downloaded on May 31,2021 at 01:10:53 UTC from IEEE Xplore. Restrictions apply.2.1.2 The use artificial neural networks to findestimation programmatic:The researcher relied on dividing the data for the 60 projectin the form of (4) to 50 projects training data sets for thenetwork, and 10 projects testing data sets for the network.Tables (4) shows training sets of neural network, Tables (5)shows test group for neural networks.Table (4): training sets of neural network training set.Table (5): test group for neural networks.The first researcher, used four neural networks:1. Feed Forward Neural Networks (FFNN)2. The Cascade Neural Network (CNN)3. Elman Neural Network (ENN)4. Radial Basis Function Training Algorithm (RBFN)2.1.2.1 Feed Forward Neural Networks FFNN:After network training and testing, the results of a trainingnetwork test gives estimate to the software effort. Table (6)shows test results of the FFNN program effort test and Table(7) shows the values of the measurements used.Table (6): test results of the FFNN. NO ProjectEffort EstimationEstimator Real51525354555657585960 55.1815 62331.0092 30058.5792 4846.2691 10.8115.1767 120286.242 370136.0855 60184.7979 210981.3089 124872.9487 72Table (7): the results of the measurements used for FFNN2.1.2.2 The Cascade Neural Network CNN:After network training and testing, the results of a trainingnetwork test gives estimate to the program effort. Tables (8)shown test results of the CNN Program effort test, and Table(9) shows the results of the measurements used.Table (8): test Result of the CNN. NO ProjectEffort estimationEffort Real51525354555657585960 50.9593 62330.4601 30063.9359 487.9962 10.8132.1036 120280.8418 370137.6326 60186.1757 210990.0450 124894.8860 72Table (9): results of the measurements used for CNN. MeasureEffort Estimator using network CNNMMRERMSEBREVAF 1.846071.39280.250193.67422.1.2.3 The results of the implementation ENN:After network training and testing, the results of the networktest gives estimate to the programming effort. Table (10)shows test results of the ENN program effort test, and Table(11) shows the values of the measurements used.Table (10): test results ENN. NO projectEffort EstimatorEffort Real51 58.8 6252 355.6 30053 38.9 4854 38.3 10.855 109.8 12056 234 37057 161.5 6058 158.5 21059 1018.3 124860 114.9 72 No ProjectActual effortKLOCEAF12……50 420 190 0.4084239 48.5 0.9694… … …… … …8.4 2.2 0.8785No project Actual effort KLOC EAF 51 62 12.8 0.969452 300 66.6 0.878553 48 15 0.471454 10.8 3.5 0.878555 120 29.5 0.878556 370 50 1.203457 60 31.5 0.523158 210 38 1.034959 1248 177.9 1.210160 72 7.5 1.5867 MeasureEffort Estimator using network FFNNMMRERMSEBREVAF 1.912373.74470.255893.3855Authorized licensed use limited to: Federation University Australia. Downloaded on May 31,2021 at 01:10:53 UTC from IEEE Xplore. Restrictions apply.Table (11): results of measurements used for ENN. MeasureEffort Estimator using network ENNMMRE 1.7408RMSE 67.0649BRE 0.2447VAF 93.0551 2.1.1.4 Results of implementation RBFN:After network training and testing, the results of thenetwork test were to estimate the program effort. Table (12)shows test results of the RBFN program effort test and Table(13) shows the values of the measurements used.Table (12): test results RBEN. No ProjectEffort EstimatorEffort Real51 70.3 6252 331.8 30053 22 4854 10.9 10.855 127.8 12056 195.1 37057 211.7 6058 156.5 21059 1579.4 124860 20.9 72 Table (13): results of measurements used RBFN.From the above, it is found that the COCOMO modelwas the worst among the methods used, it has the lowestratios by the standards used.Table (14): Comparison of neural network with the results ofCOCOMO.In general, the neural networks improved the bettermeasurement average, using the ENN was the bestamong the networks, then CNN and then FFNN, but theRBFN network was the worst, when compared toCOCOMO model.In order to compare the results of these estimation, asshown in Table [15], was prepared the values of themeasurements resulting from all methods used.Four measures were used to compare the results:a. MMRE measure: It was found that ENN is the bestnetwork to be used for estimating. The effort forthis procedure is followed by the sequential CNNnetwork, RBFN and FFNN. Also, note that theCOCOMO method achieved the lowest average,making it the worst for this the measures.b. RMSE measure: The lower the ratio, the better. TheENN is the best among the networks, followed bythe RBFN network, followed by CNN and thenFFNN. The COCOMO method is also the worstamong the methods with the highest percentage.c. BRE measure: The lower the ratio of this measure,the better the network was, and the ENN networkwas the best with the lowest percentage followedby CNN, then FFNN, then RBFN, and theCOCOMO method was the worst, this scaleranges between 1-100.d. VAF measure: Where it was closest to 100.(CNN) is the best network of local networks,(FFNN), then ENN network and COCOMOmodel. The RBFN pass is the worst for thescale.Table (15): Comparison of measure of estimation in theNeural Network with the results of COCOMO.3.1 Genetic Algorithm:Genetic Algorithms have several methods to solve aproblem, two methods of them are used in this work:x Multi -Expression Programming (MEP) MeasureEffort Estimator using network RBFNMMRERMSEBREVAF 1.898871.31930.275786.5640Mechanism of estimationMeasureRBFN ENN CNN FFNN COCOMO MMRE 1.8988 1.7408 1.8460 1.9123 2.5517RMSE71.31967.064971.392873.744136.675BRE 0.2757 0.2447 0.2501 0.2558 0.3847VAF86.56493.055193.674293.38588.9870 ActualEffortMechanism of EstimatorRBFN ENN CNN FFNN COCOMONOProject51 70.3 58.8 50.9593 55.1815 57.8513525354555657585960 62 331.8 355.6 330.460 331.0092 379.3830300 22 38.9 63.9359 58.5792 34.029648 10.9 38.3 7.9962 46.2691 11.060710.8 127.8 109.8 132.103 115.1767 142.7862120 195.1 234 280.841 286.242 234.1576370 211.7 161.5 137.632 136.0855 91.984760 156.5 158.5 186.175 184.7979 182.5637210 1579.4 1018.3 990.045 981.3089 892.66171248 20.9 114.9 94.8860 72.9487 42.1172Authorized licensed use limited to: Federation University Australia. Downloaded on May 31,2021 at 01:10:53 UTC from IEEE Xplore. Restrictions apply.x Gene-Expression Programming (GEP)3.1.1 Multi-Expression Programming MEP:Multi-Expression programming is a series of genes.The number of expressions is constant for eachchromosome. This number is known as the length of thechromosome. Each of these genes encodes either afunction or a variable. The conditions of therepresentation of multi-expression programming are:x The gene that encodes a function must contain theinput values of that function,x It should be the first gene in the chromosome of theset of variables always.One of the characteristics of this method is the abilityto store several solutions to the problem in thechromosome, usually the best choice chromosomeaccording to the fitness function will be choose.3.1.1.1 Comparison Results with GeneticProgramming:In this work, a comparison is done for the results ofimplementation of the multiple-expression programmingof NASA1 [3] data with the results of the researchers(Sheta, Al-Afeef), which used the same data withgenetic programming as shown in the following table.The MMRE and PRED to the equation of guessworkmentioned in the search results were as in Table(16), showing that MMER and PRED results areadvanced, then multiple-expression programmingalgorithm. Table [17] shows the estimation Equation forthe same data.Table (16): Comparison of multi-expression programmingresults with the results of genetic programming.Table (17): Estimation equations for data.3.1.1.2 Comparison results with neural networks:The multi-expression programming was used, and itsresults were compared with those obtained by the firstresearcher, which the used data set (NASA2) [1] andartificial neural networks was implemented to estimatethe programming effort.In this set of data, the results of multi-expressionprogramming were compared with the results of themethod of artificial neural networks. The results showssuperiority of multi-expression programming overneural network method for 7 projects, as shown in Table(18) in which the results appear in bold color as the bestvalues.The results were also compared with the skills that wereemployed using Artificial Neural Network technology toestimate the program effort. Table (19) shows thecomparison of results with the method of NeuralNetworks used by the research (PER NN) PerceptronNeural Network.Table [19] compares the results of the multi-expressionprogramming algorithm where the researcher dividedthe data into training and test. Results were based on thereal effort comparison of these projects with theestimator effort of the two methods. The MRE functionwas also applied and the results of the two methodswere compared. The results showed that the method ofthe multi-Neural Networks used by researchers with 9projects out of 15 and values in bold color show the bestresults.Table (18): the results of the multi-expression programmingcomparison with the results of four neural networks used bythe first researcher.Table (19) Comparison of the results of the PER network withthe multi-expression programming algorithm. ProjectIDActualEffortEstimated EffortMRE *100PER NNMEPPERMEP1. 120104.17108.4113.199.6552. 60 53.64 53.8 10.6 10.273. 18 18.69 22.46 3.83 24.784. 239 240.00 239.22 0.41 0.0935. 170 154.87 157.56 8.9 7.317 No. Dataset MMREM1 NASA1 0.146 0.1 NASA1 E = 2*DL – 0.59 *10PRED(25)EP143 72.2 83.3E =(sqrt(sqrt(ME)))– sqrt (sqrt(sqrt(ME) ))No DatasetEffort EquationGP(MMRE=0.146)MEP ( MMRE=0.143) GP GP MEP-3 ME2 *DLrt(ME))+DL)/(sqrt(sq ProjectIDActualEffort1. 622. 3003. 484. 10.85. 1206. 3707. 608. 2109. 124810. 72Estimated EffortANNMEPFFNNCNNENNRBFN55.1850.9558.870.359.09331.00330.46355.6331.8270.3258.5763.9338.92245.6846.267.9938.310.915.93115.17132.10109.8127.8123.22286.24280.84234195.1319.72136.08137.63161.5211.786.75184.79186.17158.5156.5190981.30990.041018.31579.4985.8272.9494.88114.920.954.25 Authorized licensed use limited to: Federation University Australia. Downloaded on May 31,2021 at 01:10:53 UTC from IEEE Xplore. Restrictions apply. 6.480439.17393.98.517.97.300245.39233.518.222.18.1817.7922.41.1624.789.5038.4240.7423.1618.510.210194.641907.319.52 3.1.2 Gene-Expression Programming GEP:The algorithm to be used is the programming of geneexpression from the main interface of the proposedinstrument, the choice of the required data, the fitnessfunction, the size of the community and the number ofgenerations. The results of the implementation of theproposed tool using the genetic expression programmingalgorithm.3.1.2.1 Comparison Results with GeneticProgramming:Table (20) shows a comparison of results of the geneexpression analysis program (NASA1) with the resultsof the researchers, and uses the same data set withgenetic programming. The results of the MMRE fitnessfunction and the PRED 25 function of geneticprogramming appeared on their counterpart of the geneexpression programming algorithm, shown by the darkcolor that shows the best values. Table [17] shows theestimation Equation for the same data.Table (20): Comparison of the results of gene expressionprogramming with the results of genetic programming.Table (21): Equations of the Gene Expression Programming.3.1.2.2 Comparison results with neural networks:Gene Expression Programming was compared to fourArtificial Neural Networks using NASA2 by theresearcher. Table (22) shows the comparison of theresults of Gene Expression Programming with theresults of Neural Networks. In Neural Networks, theyachieved the best in only 3 projects which is the best of10 jobs.Genetic Expression Programming was also comparedwith the results of Artificial Neural Networks (PER NN)using NASA2 [12][13]. Table (22) shows the results ofGene Expression Programming with the results of theNeural Network method. Genetic access to the bestsolution in 5 projects out of 15 in the case of effortestimation using the MRE function, dark blue show thebest result. The results were that the success of GeneExpression Programming is only in 5 projects and theother method used by the researcher in 10 projects.Table (22): Comparison of results of gene expressionprogramming with neural network methods.4.1 The comparison results of the equation of theCOCOMO with the Multi-Expression ProgrammingAlgorithm and the Gene Expression ProgrammingAlgorithm:As the COCOMO equation was the worst in guessingusing Artificial Neural Network, there is no need tocompare it.5.1 Comparison of Multi-Expression ProgrammingAlgorithms and Gene Expression:-Table (24) presents the results of the comparisonbetween the Multi-Expression Programming and theGene Expression performed on the data using theMMRE fitness function and the PRED function (25),and it shows the success of Multi-Expression No. DatasetMMREPRED(25)GPGEPGPGEP1. 0.146 0.18 72.2 NASA1 55.5 No Dataset1 NASA1Effort EquationGP(MMRE=0.146)GEP ( MMRE=0.18)E = 2*DL – 0.59 *10-3 ME2 *DLE=(DL/Log10(sqrt(ME))) ProjectID567ActualEffort174830Estimated Effort154.87 188.15 8.9 10.68…0GEP6.614.1401.2.3.4.5.6.7.8.9.10.8.9.10621855.18150.95.7958.821.2170.31.1660.05917.867300 331.00 330.46 355.6 331.850 38.42 40.16 23.16256.5519.648 58.57 63.93 38.9 22. 210 194.64 146.99 7.3155.95630.0010.8 46.26 7.99 38.3 10.913.482120 115.17 132.10 109.8 127.8113.63370 286.24 280.84 234 195.1208.1860 136.08 137.63 161.5 211.7117.50210 184.79 186.17 158.5 156.5155.901248 981.30 990.04 1018.3 1579.4816.1972 72.94 94.88 114.9 20.929.282 ANNFFNN CNN ENN RBFN 4. 239 240.00 279.90 0.41 17.11439.17 447.88 8.5 0 245.39 257.56 18.2 ID1.2.3.AcEtualfort143Est154.87imated Eff188.15ort8.95.6.7.7010.68A439.17NN447.88808.56.6FF00N2CNN5.39ENN257.56RBFN18.2GEP14.14462155.81850.95.7958.821.2170.31.1660.05917.868.1739.005331.000 330.46 355.6 331.8 256.5538.42 40.16 23.16 19.648258.5107 63.93 38.9 22 55.956194.64 146.99 7.31 30.0010.ProjectIDActualEffortEstimated EffortMRE *100PER NNGEPPERGEP1.120104.17114.0513.194.952.6053.6454.0910.69.83.1818.6921.213.8317.864.239240.00279.900.4117.11 Table (23): Comparison PER network with gene expressionprogramming algorithm.Authorized licensed use limited to: Federation University Australia. Downloaded on May 31,2021 at 01:10:53 UTC from IEEE Xplore. Restrictions apply.Programming for all data with respect to the MMREfunction and its success also with the PRED.Table (24): Comparison between results algorithms for thedata used. No. DatasetMMREPRED(25)MEPGEPMEPGEP1.NASA1 0.143 0.18 83.3 55.52.Kemerer0.360.4346.646.63.Desharnais0.350.5548.0527.2 As for the NASA2 data, the Multi-ExpressionProgramming was better than the programming of theGene Expression as clearly shown in the tables (26) and(27), where Multi-Expression Programming achievedsuccess in 9 projects out of 15, while GeneticExpression programming success was by only five.5.1.1 Effect of fitness function used:To determine the effect of the fitness function on theresults obtained, a number of fitness functions (MMRE,RMSE, and MAE) were compared to a set of data todetermine which fitness function is best by studying theresults using the algorithms used. Table (25) shows thedata used.Table (25): the Data used in the research. NO123NASA1[8]18 point The name of the data set Number of total projectsKemerer[10] 15 pointsDesharnais[14] 81 points: 4 incompletepoints (38,44,66,75)Tables (26), (27) and (28) show the results of theimplementation of the different fitness functions by thealgorithms and show the obvious success of the MultiExpression Programming Algorithm as shown in bold.Table (26): Comparison between the results of the twoalgorithms for the MMRE. No Dataset EGEP1.NASA10.14 0.182.Kemerer0.36 0.343.Desharnais0.35 0.55 MMRMEP Table (27): Comparison between the results of the twoalgorithms for the RMSE function. No DatasetRMSEMEPGEP1.NASA1 11.21 15.042.Kemerer 163.29 172.44 3.Desharnais2609.142875.46 Table (28): Comparison of the results of the two algorithms ofthe MAE function. No DatasetMAEMEPGEP1.NASA1 6.77 6.832.Kemerer 88.61 90.833.Desharnais 1649.91 2209.08 6.1 The system of estimating the effort of thesoftware using the cluster with Fuzzy Logic:Intelligent techniques are used extensively in thefield of software estimation. A good estimator algorithmusing a fuzzy cluster system (FCM). It is essential forsoftware evaluators to take several basic steps, the firstis the collection of data for the purpose of reaching therequired solution, where data is collected from previouscomplete and developed projects, in this research adatabase was used which consists of NASA 60 Program.This data contains values specific to the characteristicsaffecting any software project 15 cost factors.In addition, the size of the project is measured by theKLOC as well as the real effort of the project in sizePM. After completion of the data collection process, weidentify the characteristics that will be selected as aninput to the Fuzzy system where there are severalattempts to identify these characteristics embodied inthe introduction of all data.The 15 factor costs in addition to the KLOC, orselecting the most influential factors on the softwareproject and its use as an input to this system in additionto the KLOC.After collecting the data and defining the characteristicsto be entered into the fuzzy system for the purpose ofbuilding and designing this system, a data set isfragmented, NASA 60 is used to two group sets, trainingand testing.The 50 projects are identified as training data and 10projects as test data as shown in the tables (4) and (5(.The following is an explanation of the basic steps fordesigning and building a good software estimationsystem using a fuzzy cluster FCM:Authorized licensed use limited to: Federation University Australia. Downloaded on May 31,2021 at 01:10:53 UTC from IEEE Xplore. Restrictions apply.Configuration of the buffering system inputs in theconfiguration phase of the membership functions, thedata centers and the number of outputs of the systemwhere:x Fuzzy system inputs, the EAF vector, and the KLOCvector including columns, as each row represents oneproject.x System inputs which represent the estimated effort ofthese projects.x The number of concentration centers by which thenumber of classes (nodes) is determined in the FCMalgorithm and the number of organic functions in thegiven logic.x The type of inference used in the logic of the point,where in this field the use of the type of Mamdaniderived from the type Sugeno in the process ofestimating the program effort.After several attempts to determine the best type ofinference, Mamdani or Sugeno in the Fuzzy system,aggregation system used in estimating the programmingeffort and selecting the best range for the cluster centersof each type. It was found that the Sugeno is the best inestimating effort due to its good results compared to theresults of number (2) which was chosen as the bestnumber of data centers in the heuristics type (Sugeno)and the number (10) in the Mamdani inference to solvethe problem of estimate as shown in table (29).Table (29) shows that inference the type of Sugeno wasbetter and has given the ratios of accurate measurementstype Mamdani until the increase in the number ofcenters in the inference Mamdani. Even though, itremains less accurate than Sugeno inference, where itgot the lowest percentage of measure by standards. Thebest number of cluster centers adopted by the mamdanisystem is 10, while the best number of cluster centersused by Sugeno.6.1.1 Results of the implementation of the fuzzysystem (FCM):After the formation and construction of the aggregationsystem and the testing of data, results of this system areobtained, which represents the value of the estimated programeffort required. Table (30) illustrates this matter.Table (29): Number of nodes Centers and Type of InferenceUsed in the fuzzy C-means System (FCM).Table (30) compare the results Effort estimator by FCM(Mamdani , Sugeno) and effort real. Effort estimatorby FCM SugenoEffort estimatorby FCM MamdaniEffortRealNOproject51 62 166.1192 6.790252 300 409.1309 331.731653 48 148.4179 180.517454 10.8 150.4162 73.327155 120 216.9768 93.575056 370 274.4157 369.634157 60 169.2488 51.610958 210 213.3883 51.610959 1248 1077.3 1193.660 72 159.4271 267.2433 Table (30) shows results of the two components usingthe Fuzzy aggregation system FCM Mamdani, FCMSugeno. Once the calculated effort values have beenobtained, the efficiency measures used are calculated,the table shown (31).No Numberof centersType ofinferenceMeasuresMMRE RMSE BRE VAF1.2Mamdani 14.234 2738.2 9.954 38.18Sugeno 2.7117 9.9547 0.579 95.432.3Mamdani 5.478 819.45 4.231 32.64Sugeno 2.7117 96.434 0.579 95.433. 4 Mamdani 2.3651 277.68 4.231 59.02Sugeno 2.7117 96.434 0.579 95.434. 5 Mamdani 2.0862 265.75 2.303 63.73Sugeno 2.7117 96.434 0.579 95.435. 6 Mamdani 1.3653 161.34 2.044 77.85 Sugeno 2.7117 96.434 0.579Mamdani 1.3601 167.79 1.695 95.436. 7 73.83Sugeno 2.7117 96.434 0.579 95.437. 8 Mamdani 1.2512 193.28 1.482 89.62Sugeno 2.7117 96.434 0.579 95.438. 9 Mamdani 1.2629 138.19 1.340Sugeno 2.7117 96.434 0.579 89.6295.439. 10 Mamdani 1.2143 226.46 1.322 89.39Sugeno 2.7117 96.484 0.579 95.4310. 11 Mamdani 1.4046 230.48 1.339 92.30Sugeno 2.7117 96.434 0.579 95.4311. 12 Mamdani 2.0203 249.92 1.147 93.35 Sugeno 2.7117 96.434 0.579Mamdani 1.9581 249.92 1.147 95.4312. 13 90.78Sugeno 2.7117 96.646 0.579 95.4313. 14 Mamdani 1.3678 449.93 1.126 65.34 Sugeno 2.7117 96.434 0.579 95.43Authorized licensed use limited to: Federation University Australia. Downloaded on May 31,2021 at 01:10:53 UTC from IEEE Xplore. Restrictions apply.Table (31): Measure Effort estimator FCM Mamdani andEffort estimator by FCM Sugeno. Effort estimatorby FCM SugenoEffort estimator byFCM MamdaniMeasureMMRE 1.2143 2.7117RMSE 158.8286 96.4349BRE 1.2946 0.5970VAF 92.3032 95.3406 6.2 Use the NARX neural network to estimate theprogrammable effort:After the NARX network was built, validated andtested, the results of this network were obtained, whichrepresent the value of the required programming effort,the table shown (32).Table (32) compare the results effort real and result ofNARX. Effort estimator bynetwork NARXNO projectEffort Real516248.658252300300.8216534877.73295410.84.988855120138.866056370303.54415760149.90535821019957345912481162.2607251.9701 The values of the network efficiency measures usedwere calculated. Table (33) illustrates this matter.Table (33): Measure Effort estimator of NARX.6.3 Results of methods used and compared above:Some of the techniques were used in estimating theprogramming effort, which included the Fuzzy systemFCM and the NARX. After the construction andimplementation of these techniques, the final resultswere obtained for estimating the program effort. Table(34) illustrates the matter.Table (34): Effort estimator FCM (Mamdani , Sugeno) andNARX compare of effort real. The estimate mechanism usedEffortRealNOprojectEffortestimatorEffortestimatorEffortestimator by by networkNARXby FCMSugenoFCMMamdani51 62 166.1192 26.7902 48.658252 300 409.1309 331.7316 300.821653 48 148.4179 180.5174 77.732954 10.8 150.4162 73.3271 4.988855 120 216.9768 93.5750 138.866056 370 274.4157 369.6341 303.544157 60 169.2488 51.6109 149.905358 210 231.3883 217.6807 199573459 1248 1077.3 1193.6 1162.260 72 159.4271 267.2433 51.9701In order to compare the results of these estimates, Table(35) was prepared with the values of the resultingmeasurements for the methods used.Table (35): Measure Effort estimator FCM (Mamdani,Sugeno) and NARX.Effort estimatorby networkNARXEffortestimator byFCM SugenoEffortestimator byFCM MamdaniMeasureMMRE 1.2143 2.7117 0.2782RMSE 158.8286 96.4349 19.7811BRE 1.2946 0.5970 0.1498VAF 92.3032 95.3406 98.26107.1 Conclusion:1. The results in Table (36) show that Neural Networksare generally the best of the three methods, where theirresults have achieved better convergence with the actualresults. Competition remains among the types of neuralnetworks (FFNN, CNN, ENN, RBFN, NARX), wherewe note that the NARX network was the best network.2. The method of the Genetic Algorithm proved thatMulti MEP Expression Programmin
