EVADING: An Evolutionary Algorithm with Dynamic Niching for Data Classification.
John Psaroudakis, Fani A. Tzima, Pericles A. Mitkas · GEM · 2009
Multimodal optimization problems (MMOPs) have been widely studied in many fields of machine learning, including pattern recognition and data classification. Formulating the process of rule induction for the latter task as a MMOP and inspired by corresponding findings in the field of function optimization, our current work proposes an evolutionary algorithm (EVADING) capable of discovering a set of accurate and diverse classification rules. The proposed algorithm uses a dynamic clustering technique as a parallel niching method to maintain rule population diversity and converge to the optimal rules for the attribute-space defined by the target dataset. To demonstrate its applicability and potential, EVADING is applied to a series of real-life classification problems and its prediction accuracy is compared to that of other popular non-evolutionary machine learning techniques. Results are encouraging, since EVADING manages to achieve the best overall average ranking and performs significantly better (at significance level a=0.05) from three out of the eight rival algorithms used in this study. This work is concluded with some insights on the factors affecting the proposed algorithm’s performance, along with the directions of our future research.