Hybrid Computational Models for Software Cost Prediction: An Approach Using Artificial Neural Networks and Genetic Algorithms

Abstract

Over the years, software cost estimation through sizing has led to the development of various estimating practices. Despite the uniqueness and unpredictability of the software processes, people involved in project resource management have always been striving for acquiring reliable and accurate software cost estimations. The difficulty of finding a concise set of factors affecting productivity is amplified due to the dependence on the nature of products, the people working on the project and the cultural environment in which software is built and thus effort estimations are still considered a challenge. This paper aims to provide size- and effort-based cost estimations required for the development of new software projects utilising data obtained from previously completed projects. The modelling approach employs different Artificial Neural Network (ANN) topologies and input/output schemes selected heuristically, which target at capturing the dynamics of cost behavior as this is expressed by the available data attributes. The ANNs are enhanced by a Genetic Algorithm (GA) whose role is to evolve the network architectures (both input and internal hidden layers) by reducing the Mean Relative Error (MRE) produced by the output results of each network.