Duplicate Individuals in Differential Evolution
Tomofumi Kitamura, Alex S. Fukunaga · 2022 IEEE Congress on Evolutionary Computation (CEC) · 2022
Some modern computing architectures such as those which use GPUs offer significantly more computational capability in single-precision than double-precision floating point. However, when using single-precision math, evolutionary optimization algorithms such as differential evolution are much more prone to collision, where new individuals created by reproduction operators are duplicates of previously generated individuals. We show experimentally that on the CEC2014 single objective benchmarks and the BBOB benchmarks, search using the SHADE adaptive DE can waste more than 20% of the fitness evaluations due to collisions. We (1) discuss the causes of this collision, (2) detect collisions using hash to avoid unnecessary evaluations, and (3) propose a restart method based on collisions.