Experimental Evaluation of Adaptive Operators Selection Methods for the Dynamic Multiobjective Evolutionary Algorithm Based on Decomposition (DMOEA/D)

Abstract

This paper compares three methods of adaptive selection of operators (AOS) incorporated into the Dynamic Multiobjective Evolutionary Algorithm Based on Decomposition (DMOEA/D). The first method is structured in several continuous sections and is designed to execute a test and application procedure to find the way that allows a better adaptive selection of the best available operators. The second method is based on a multi-armed bandit problem in which Thomson Dynamic Sampling (DYTS) is applied. Each arm represents a reproduction operator and is assigned a previous reward distribution. This method chooses an operator through sampling that rewards distribution according to DYTS. Finally, the third method also focused on multi-arm bandits and adding the range of the fitness rate to track the dynamics of the process and uses a sliding window to record the rates of improvement of the current fitness achieved by the operators, at the same time applying a decrementing mechanism to increase the probability of selecting the best operator. The computational experiments evaluated four dynamic reference problems called FDA1, FDA3, dMOP1, and dMOP2 and applied hypervolume metrics, generalized dispersion, and inverted generational distance. For the feasibility validation of these methods, the Wilcoxon and Friedman non-parametric tests were applied with a significance level of 5%.