Adaptation in differential evolution: A numerical comparison

Abstract

Paper considers adaptation of control parameters in differential evolution. Adaptation by competitive setting is described and two novel variants of competitive differential evolution are proposed. Five adaptive variants of differential evolution are compared with other search algorithms on three benchmarks. One of them is the novel composition test functions, where the variants of differential evolution outperform other algorithms in 5 of 6 test functions. The NIST nonlinear regression datasets are used as the second benchmark and a subset of CEC’05 benchmark functions as the third one. The performance of adaptive differential evolution is compared with the adaptive controlled random search algorithm, tailored especially for the nonlinear-regression problems. Two of five tested variants of adaptive differential evolution are almost as reliable as the adaptive controlled random search algorithm and one of these variants converges only slightly slower than the adaptive controlled random search in nonlinear-regression problems. The results achieved in CEC’05 benchmark functions are close to the best performing algorithm. Therefore, the adaptive differential evolution is a promising tool of heuristic search for the global minimum in boundary-constrained problems.