Proceedings of SAI Intelligent Systems Conference (IntelliSys) 2016

Abstract

In software development, project professionals usually rely upon their preceding experience as a way to estimate the quantity of men/hours for cost estimation. Software products are acceptable by clients as long as they are developed within the budget. Therefore, accurate prediction of software development cost is an extremely important phase before starting the actual development phase. Practitioners, who are about accurate prediction, admit their own inability of estimating development cost. In recent literature, a number of Machine Learning (ML) based techniques have been proposed for accurate prediction of software costs. The main objective of this paper is review, analyse, and critically evaluate ML techniques utilized for cost estimation and identify the limitations in the existing techniques. It has been identified that ML-based techniques have been successfully employed for accurate cost estimation but a number of issues remained unresolved in the prior literature. Firstly, the employed techniques have been tested with traditional benchmark datasets which reflect the use of conventional development methodologies, that is, Waterfall and the evidence provided in research could not be used for projects which are developed using new development methodologies such as Incremental or Agile as such newly developed project data is much richer in information as compared to the traditional project related data. Secondly, previously proposed models have not been evaluated thoroughly using advanced evaluation measures. There is a strong need of a revised ML-based model for accurate cost estimation which not only utilizes the rich information present in the projects developed using new methodologies but also provides wider applicability. We propose a new model which exploits multilayer perceptron technique with effective feature selection methods for improving software cost estimation. The proposed model has been validated using multiple real-world datasets.