Over the past two decades, reinforcement learning has yielded phenomenal successes in the domain of perfect-information games: it has produced world-class bots capable of outperforming even the strongest human competitors in games such as Chess and GO (Silver et al., 2018). Most exciting of all, the resulting poker bot is highly interpretable, allowing humans to learn from the novel strategies it discovers.
Through experiments against Slumbot, the winner of the most recent Annual Computer Poker Competition, we demonstrate that our approach yields a HUNL Poker agent that is capable of beating the Slumbot. Second, we make use of globally optimal decision trees, paired with a counterfactual regret minimization (CFR) self-play algorithm, to train our poker bot which produces an entirely interpretable agent.
Namely, first we propose a novel, compact, and easy-to-understand game-state feature representation for Heads-up No-limit (HUNL) Poker. In this paper, we present advances on both fronts. This lack of interpretability has two main sources: first, the use of an uninterpretable feature representation, and second, the use of black box methods such as neural networks, for the fitting procedure. We address the problem of interpretability in iterative game solving for imperfect-information games such as poker.
