Exploring self-play as an alternative to coevolution

ABSTRACT This paper introduces a novel variant of self-play training which is used for optimizing functions whose size of parameter space computationally precludes the sustainedapplication ofconventional population coevolution for a large number of iterations. The proposed algorithm provides an effective way to fine-tune and adjust function parameters towards their optimal values. I use chess as a test bed but the underlying algorithm is applicable wherever self-competition can be effectively defined. The randomly initialized organisms undergo three phases of training (1) using conventional evolutionary algorithms (2) training using self-play (3) single population coevolution. The algorithm is intended to be used to provide similar performances as the SOTA but at a much less computational expenditure.