Optimizing Non-orthogonal Space Distance Using PSO in Software Cost Estimation

This paper proposes a method to optimize the Nonorthogonal Space Distance (NoSD) based on the Particle Swarm Optimization (PSO) algorithm so as to increase estimation accuracy in analogy-based software cost estimation. NoSD is a measure of projects similarity that uses a matrix defined based on mutual information to take both feature redundancies and feature weights into distance computation. We assumes that such definition based only on mutual information between features can hardly describe real-life software projects accurately, so we proposes this new method and improves NoSD using optimization techniques. In this proposed method, the matrix in NoSD is optimized by the PSO algorithm with the goal of minimizing estimation error at training stage. Based on this optimized matrix, which better fits real-life software projects, the distance definition can measure projects similarity more accurately and thus can greatly improve the estimation accuracy. Experiments have been conducted on two real-life software projects datasets (Desharnais and ISBSG R8) using the proposed method along with several other widely used methods including Euclidean, Manhattan, Minkowski, Mahalanobis, NoSD, and weighted Euclidean distance. Results show that this method brings notable improvements in estimation accuracy based on three widely used evaluation metrics: MMRE, MdMRE, and PRED(0.25).