Adaptive epsilon-Sampling and epsilon-Hood for Evolutionary Many-Objective Optimization

Abstract

Many-objective problems are becoming common in several real-world application 
domains and there is a growing interest to develop evolutionary many-objective 
optimizers that can solve them effectively. Studies on selection for many-objective 
optimization and most recently studies on the characteristics of many-objective 
landscapes, the effectiveness of operators of variation, and the effects of large 
populations have proved successful to advance our understanding of evolutionary 
many-objective optimization. This work proposes an evolutionary many-objective 
optimization algorithm that uses adaptive epsilon-dominance principles to select 
survivors and also to create neighborhoods to bias mating, so that solutions will 
recombine with other solutions located close by in objective space. We investigate 
the performance of the proposed algorithm on DTLZ continuous problems, using a 
short number of generations to evolve the population, varying population size from 
100 to 20000 individuals. Results show that the application of adaptive epsilon-dominance 
principles for survival selection as well as for mating selection improves considerably the 
performance of the optimizer.