Connect6 Opening Leveraging AlphaZero and Job-Level Computing

Abstract

For most board games, players commonly learn to increase strengths by following the opening moves played by experts, usually in the first stage of playing. In the past, many efforts have been made to use game-specific knowledge to construct opening books. Recently, DeepMind developed AlphaZero (2017) that can master game playing without domain knowledge. In this paper, we present an approach based on AlphaZero to constructing an opening book. To demonstrate the approach, we use a Connect6 program trained based on AlphaZero for evaluating positions, and then expand the opening game tree based on a job-level computing algorithm, called JL-UCT (job-level Upper Confidence Tree), developed by Wu et al. (2013) and Wei et al. (2015). In our experiment, the strengths of the Connect6 programs using this opening book are significantly improved, namely, the one with the opening book has a win rate of 65% against the one without using the book. In addition, the one without opening lost to Polygames in the Connect6 tournament of TCGA 2020 competitions, while the one with opening won against Polygames in TAAI and Computer Olympiad competitions later in 2020.