Design and analysis of machine learning models for software effort estimation

Abstract

Software Effort Estimation is one of the challenging tasks in software project management. The main purpose of effort estimation is to accurately estimate the resources and schedule the projects with reference to time and environment. It is usually done at the early stage of project management by Project Managers (PM). The project managers estimate the effort using various algorithmic strategies such as Analogy based estimation, Expert judgement estimation etc. It is found that many machine learning algorithms have been used to solve various problems in software engineering. Moreover, some machine learning algorithms have been already applied to estimate total software development effort in the last few years. However, the performance of the existing methods is yet to be improved. This research work aims to improve the prediction accuracy of effort estimation using machine learning methods. It also uses ensemble machine learning model to improve the accuracy of estimation. It proposes three different methods to compute the effort. In the first method, Fuzzy Logic model is used to estimate the software development effort. The main reason to use fuzzy logic is to eliminate the unambiguity and uncertainty in the data set. Unlike the existing fuzzy based approach, the proposed method uses the Trapezoidal Membership Function (TMF) for fuzzification of values in the data set. The proposed approach uses multiple IF-THEN rules on the fuzzy values. When the resultant fuzzy value is converted The accuracy of the proposed model is evaluated using the performance metrics Mean Relative Error (MRE). This model produces low error rate when compared to COCOMO Model. Although the first method improves the performance, the prediction accuracy, it requires more number of IF-Then rules. As regression methodologies such as Least Square Regression (LSR), Adaptive REcursive data partitIONing (AREION), Quick method are widely used for prediction in many domains, these have been applied in this research to estimate effort accurately. The proposed methods are tested on the data sets like Desharnais, Cocomo81, CocomoNasa63 and CocomoNasa90 are collected from publically available repository. The performance analysis of the proposed methods is done by using the metrics like Mean Relative Error (MRE), Magnitude of Error Relative (MER), Mean Magnitude of Error Relative (MMER), Median of MER (MdMER). The experimental results show that ARIEON methodology with log transformation produces accurate results when compared to other methodologies. However the proposed method marginally improves the performance compared to the first approach. Therefore a new hybrid model is proposed to compute the total effort. newline