A machine learning based approach for cost and effort estimation in agile development process

Abstract

In current scenario of software industry culture, an important and crucial task under project management is accurate estimation of practical measures like cost and effort which subsequently results in successful project completion. Many researchers have analysed and proposed various techniques in the estimation for software projects using conventional frameworks like waterfall, incremental etc. In recent years as there are technological advancements and there is a requirement of adaptation to technological changes, hence agile development methodology has attracted the interest of many researchers and software developers in software companies. Various researchers have proposed several techniques including opinion based, algorithm based and machine learning based techniques for effort and cost estimation of software projects. The proposed work in this study deals with the study and analysis of the most popular techniques used in every category of estimation practices used in agile development. Further, the machine learning techniques based on regression, and few ensemble machine learning techniques like Ada-Boost Regressor, Bagging Regressor and Extra Trees Regressor are applied and the results are observed and analysed for the estimation accuracy. Based on all the observations the work proposes a novel technique (SVC_NN Integrated Regressor) using ensemble of SVR along with the RBF kernel and ANN backpropogation for the effort estimation of software projects. The proposed technique customizes the Support Vector Regressor with RBF kernel along with 2 other factors, that is, the slope and the biasing constant. By experimental validations, it is found that the prediction trend of the proposed model lacks these parameters which are key components to all linear models. Hence, to get the most optimized result for the proposed integrated model, we have used a small perceptron model to predict the slope and intercept constant value for the model. The work also provides a comparison of the proposed technique with the resul