Extracting tactics learned from self-play in general games

Local, spatial state-action features can be used to effectively train linear policies from self-play in a wide variety of board games. Such play games directly, or bias tree search agents. However, the resulting feature sets large, with significant amount overlap and redundancies between features. This is problem for two reasons. Firstly, large computationally expensive, which reduces playing strength agents based on them. Secondly, correlations impair ability humans analyse, interpret, understand tactics learned by policies. We look towards decision trees their perform selection, serve as interpretable models. Previous work distilling into uses states inputs, distributions over complete action space outputs. In contrast, we propose evaluate types, take pairs provide various different types outputs per-action basis. An empirical evaluation 43 presented, those are case studies where attempt interpret discovered