A performance assessment of modern Niching methods for parameter optimization problems

Niching genetic algorithms (NGAs) are designed to locate multiple fitness function optima. Numerous NGAs exist: but an accurate picture of their relative strengths and weaknesses remains elusive. The variety of performance measures and experimental methodologies makes accurate comparison difficult. Test functions are also limited in number, and possess structural regularities. Furthermore, most NGAs require determination of one or more parameters, but the issue of parameter sensitivity is rarely explored. Here, we study the performance and sensitivity of several NGAs using a common experimental methodology. We consider several new nonuniform test functions, in addition to those commonly used. Finally, NGA researchers have almost exclusively used binary variable encodings. We also analyze NGA performance under both binary and gray encodings.