Evaluation of Differential Evolution Algorithm with Various Mutation Strategies for Clustering Problems

Abstract

Evolutionary Algorithms (EAs) based pattern recognition has emerged as an alternative solution to data analysis problems to enhance the efficiency and accuracy of mining processes. Differential Evolution (DE) is one rival and powerful instance of EAs, and DE has been successfully used for cluster analysis in recent years. Mutation strategy, one of the main processes of DE, uses scaled differences of individuals that are chosen randomly from the population to generate a mutant (trial) vector. The achievement of the DE algorithm for solving optimization problems highly relies on an adopted mutation strategy. In this paper, an empirical study was presented to investigate the effectiveness of six frequently used mutation strategies for solving clustering problems. The experimental tests were conducted on the most widely used data set for EAs based clustering, and the quality of cluster solutions and convergence characteristics of DE variants were evaluated. The obtained results pointed out that the mutation strategies that use the guidance information from the best solution mange to find more stable results whereas the random mutation strategies are able to find high quality solutions with slower convergence rate. This study aims to provide some information and insights to develop better DE mutation schemes for clustering.