On the utility of randomly generated functions for performance evaluation of evolutionary algorithms

Previous researches have disclosed that the excellent performance of some evolutionary algorithms (EAs) highly depends on existence of some properties in the structure of the objective function. Unlike classical benchmark functions, randomly generated multimodal functions do not have any of these properties. Having been improved, a function generator is utilized to generate a number of six benchmarks with random structure. Performance of some EAs is evaluated on these functions and compared to that evaluated on results from classical benchmarks, which are available in literature. The comparison reveals a considerable drop in the performance, even though some of these methods have all possible invariances. This demonstrates that in addition to properties, classical benchmarks have special patterns which may be exploited by EAs. Unlike properties, these patterns are not eliminated under linear transformation of the coordinates or the objective function; hence, limitations should be considered while generalizing performance of EAs on classical benchmarks to practical problems, where these properties or patterns do not necessarily exist.