Tournament-based convection selection in evolutionary algorithms

Abstract

One of the problems that single-threaded (non-parallel) evolutionary algorithms encounter is premature convergence and the lack of diversity in the population. To counteract this problem and improve the performance of evolutionary algorithms in terms of the quality of optimized solutions, a new subpopulation-based selection scheme – the convection selection – is introduced and analyzed in this work. This new selection scheme is compared against traditional selection of individuals in a single-population evolutionary processes. The experimental results indicate that the use of subpopulations with fitness-based assignment of individuals yields better results than both random assignment and a traditional, non-parallel evolutionary architecture.