Vanilla and Monte Carlo Counterfactual Regret Minimization for Imperfect Information Games

Over the past decade, Artificial Intelligence research in imperfect information games has witnessed remarkable progress, culminating in superhuman performance in games as large as No-Limit Hold’em. This advancement not only revolutionizes the realm of games, but also holds profound implications for real-world applications where strategic interactions involve multiple participants and hidden information, such as financial markets, negotiations, and security interactions. The recent development of Poker Artificial Intelligence systems like Cepheus, Libratus, and Rebel marks significant milestones in AI’s ability to tackle complex strategic environments characterized by imperfect information and even multiple participants. This project aims to explore imperfect information games with the final goal of training agents to play Kuhn Poker and Leduc Hold’em at an unexploitable level.