Bluff and Learn

We present the first systematic AI framework for playing Liar Bar, a challenging multi-player bluffing game with imperfect information, sequential deception, and stochastic penalties. By implementing and comparing Counterfactual Regret Minimization (CFR) and Neural Fictitious Self-Play (NFSP) across both two- and four-player game modes, we highlight the critical trade-offs between tabular and neural approaches to equilibrium learning. Our results show that while CFR performs well in simplified settings, NFSP—enhanced with Double DQN and dense per-round rewards—achieves superior generalization, faster convergence, and near-zero exploitability in complex scenarios. This work not only introduces the first AI agents for Liar Bar but also provides a robust benchmark for future research on scalable, strategy-aware learning in high-risk, imperfect-information environments.