An empirical investigation of evolutionary algorithm for software maintainability prediction

Abstract

Software maintenance is one of the tedious as well as costly phases in the software development life cycle. It starts immediately after the software product is delivered to the customer and ends when the product is no longer in use. There are various activities carried out during software maintenance phase such as the addition of new features, deletion of obsolete features, correction of errors, adaption to new environment etc. Software maintainability is the quality attribute of the software product which determines the ease with which these modifications can be performed. If we can predict the maintainability accurately, cost and time associated with the maintenance activity can be highly reduced. The main aim of this study is to propose the use of evolutionary technique particularly genetic algorithm for the software maintainability prediction and compare its performance with various machine leaning techniques such as Decision Table, Radial Basis Function Neural Network, Bayes Net and Sequential Minimal Optimization (SMO). In order to carry out this empirical investigation, datasets from four open source software systems are collected. The maintenance effort is calculated by counting the number of changes in terms of line of code from one version of the software to another. Based on the experiments conducted, we conclude that the evolutionary algorithm outperformed all the other classifiers, thus, very useful for the concise prediction of software maintainability. Results of this would be helpful to practitioners as they can use the maintainability prediction in order to achieve precise planning of resource allocation.