Rational models to improve performance of differential evolution for MOEA/D
Chixin Xiao, Zhigang Xue, Chixin Xiao, Jianping Yin · 2014
According to the response surface methodology (RSM), this paper proposes a new trial vector perturbation model for Differential Evolution(DE). It not only consists of linear and nonlinear heuristics for optimum exploring, but also retains simplicity style of DE mutation operator in appearance which can help the new algorithm absorb many existed elegant methods from a large amount of DE variants. The new DE-like algorithm equipped with two well-designed approximation models is faster and more reliable to search optimum in a relatively small region than just to do it directly via the original fitness function, especially, when the decision space have many local optima or the optimum are located in a very narrow space. In order to examine its performance completely, the proposed model is developed further to be embedded into MOEA/D frame to solve some challenging multi-objective benchmarks. The promising test results show rational information can be very useful, as long as to be applied properly, rather than be the patent of premature convergence.