CoupVisor: AI Decision Support System for Coup Card Game

CoupVisor is a decision-support tool for the concealed-information card game Coup, detailed in a paper on arXiv (2608.15868). It tackles two primary inquiries: the optimal action for a player each turn and the appropriate timing to contest an opponent's assertion. The system operates on a cohesive framework of game events, facilitating manual gameplay, the replay of recorded matches, simulations, belief tracking, recommendations from advisors, and policies based on learning. CoupVisor calculates the likelihood of a claim's veracity by integrating the probabilities of each role with the claimant's card count, addressing a flaw where initial claims were unjustly deemed suspicious. The research contrasts a rule-based advisor with various learned and heuristic players in numerous simulated games against diverse opponent strategies. A key insight is that the type of reward—favoring immediate rewards versus overall victory—shapes the efficacy of the learning method. This system aims to offer strategic insights in a game characterized by deception and deduction, presenting an innovative method for decision support in imperfect-information scenarios.