Estimation of effort and costs in the development of software projects using artificial neural networks based on Taguchi’s orthogonal vector plans

Abstract

The modern software industry requires fast, highquality, and accurate forecasting of efforts and costs before the actual effort is invested in realizing the software product. Such requirements are a challenge for any software company, which must be ready to meet the expectations of the software customer. The main factor in the successful development of software projects and reducing the risk of errors is an adequate of the effort and costs invested during its implementation. In this doctoral dissertation, different approaches and models that have not been sufficiently precise and efficient so far will be analyzed, which resulted in only about 30% of successfully implemented software solutions. The main goal is to present three new, improved models based on efficient artificial intelligence tools, artificial neural networks. All three improved models use different architectures of artificial neural networks (ANN), constructed based on Taguchi's orthogonal vector plans. The goal is to optimize the improved models to avoid repeating the number of experiments and long time for their training. Applying the clustering method to several different sets of real projects further mitigates their heterogeneous structure. In addition, the input values of the projects are homogenized by the method of fuzzification, which achieves even greater reliability and accuracy of the obtained results. Optimization by the Taguchi method and increasing the coverage of a wide range of different projects leads to the efficient and successful completion of as many different software projects as possible. The main contributions of this paper are: constructing and identifying the best model for estimating effort and cost, selecting the best ANN architecture whose values converge the fastest to the minimum magnitude relative error, achieving a small number of experiments, reduced software effort estimation time due to convergence rate. Additional criteria and constraints are introduced to monitor and execute experiments using a precise algorithm to execute all three new proposed models. In addition to monitoring the convergence rate of each artificial neural network architecture, the influence of the input values of each model on the change in the value of the magnitude relative error of the model is also monitored. The models constructed in this way have been experimentally checked and confirmed several times on different sets of real projects and can be practically applied, and the obtained results indicate that the achieved error values are lower than those presented so far. Therefore, the proposed models in this dissertation can be reliably applied and used to assess the efforts and costs for software development and projects in other areas of industry and science.

Savremena softverska industrija zahteva brzo, kvalitetno i precizno predviđanje napora i troškova, pre nego što se stvarni napor uloži u realizaciju softverskog proizvoda. Ovako postavljeni zahtevi predstavljaju izazov za bilokoju softversku kompaniju, koja mora biti spremna da ispuni postavljenaočekivanja naručioca softvera. Glavni faktor uspešnog razvoja softverskihprojekata i smanjenja rizika od grešaka je adekvatna procena uloženog naporai troškova tokom njegove realizacije. U ovoj doktorskoj disertaciji biće analizirani dosadašnji različitih pristupi i modeli koji nisu u najvećoj meri bilidovoljno precizni i efikasni, što je za posledicu imalo samo oko 30% uspešnorealizovanih softverskih rešenja.Glavni cilj biće predstavljanje tri nova poboljšana modela zasnovana na efikasnim alatima veštačke inteligencije, veštačkim neuronskim mrežama (engl. Artificial neural networks/ANN). Sva tri poboljšana modela koristerazličite arhitekture veštačkih neuronskih mreža, konstruisanih na osnovuTagučijevih ortogonalnih vektorskih planova. Cilj je optimizacija poboljšanihmodela kako bi se izbeglo ponavljanje broja eksperimenata i dugotrajno vremeza njihovo obučavanje, odnosno treniranje. Primenom metode klasterizacije naviše različitih skupova realnih projekata dodatno se ublažava njihova heterogena struktura. Dodatno, ulazne vrednosti projekata se homogenizuju metodom fazifikacije čime se postiže još veća pouzdanost i tačnost dobijenih rezultata. Optimizacija Tagučijevom metodom uz povećanje pokrivenosti širokog spektra različitih projekata, dovodi do efikasnog i uspešnog dovršavanja što više različitih softverskih projekta. Glavni doprinosi ove disertacije su: konstruisanje i identifikovanje najboljeg modela za procenu napora i troškova, odabir najbolje ANN arhitekture čije vrednosti najbrže konvergiraju minimalnoj magnitudnoj relativnoj greški, postizanje malog broja izvedenih eksperimenata, smanjeno vreme procene softverskog napora zbog stope konvergencije. Uvode se i dodatni kriterijumi i ograničenja za nadgledanje i izvršavanje eksperimenata pomoću preciznog algoritma za izvršavanje nad sva tri nova predložena modela. Pored praćenja brzine konvergencije svake arhitekture veštačke neuronske mreže, prati se i uticaj ulaznih veličina svakog modela na promenu vrednosti magnitudne relativne greške modela. Na ovakav način konstruisani modeli eksperimentalno su više puta proveravani i potvrđeni na različitim skupovima realnih projekata i mogu se praktično primenjivati, a dobijeni rezultati ukazuju da su postignute vrednosti grešaka niže od dosadašnjih predstavljenih. Samim tim se predloženi modeli u ovoj disertaciji mogu pouzdano primenjivati i koristiti ne samo za procenu napora i troškova za razvoj softverskih već i za razvoj projekata u drugim oblastima industrije i nauke