Solving Imperfect Information Games: Multiplayer Poker AI

Abstract

We explore the various algorithms and methods used in solving, training and constructing bots for poker games. Due to the complexity of parameters in multiplayer No-Limit Texas Hold’em and limited resources, we address much simplified versions of poker games for reasoning and implementation. Nevertheless, these simplified versions encapsulate the fundamental characteristics—incomplete information, position, action streets, randomness, etc.— that need be addressed in much larger games and can be utilized to give further insight. We implement and train poker agents based on the counterfactual regret inimization algorithm to simulate fictious play and observe that we reach a state of -Nash equilibrium despite the absence of theoretical guarantees.