Optimizing strategy selection in hidden role games

We address hidden-role decision making under uncertainty in The Resistance: Avalon. We present DeepBayes, which augments a standard Counterfactual Regret Minimization Plus (CFR+) decision procedure with two complementary inference components. First, a history-driven role assignment prediction network generates role-assignment hypotheses from past gameplay, which are used to improve the estimation of Counterfactual Values (CFVs). Second, a Bayesian Identity Recognition (BIR) method produces explicit posterior beliefs about opposing identities online as play unfolds. During CFR+ iterations, the algorithm selects actions by jointly considering the CFVs estimated under the generated role assignments and the posterior beliefs from BIR. In five-player Avalon experiments, DeepBayes achieves consistent gains in win rate over strong baselines.