Creating a Competitive Multiplayer Pokerbot Using Strategy Stitching and Online Learning

Abstract

Due to the imperfect information and stochastic elements in poker, it has proven to be a difficult task to produce computer programs that can be competitive with expert human players. In this thesis, we extend the current research on using counterfactual regret minimization in heads-up play to multiplayer poker and attempt to find an optimal tradeoff between speed and accuracy in poker abstraction by employing strategy stitching techniques to expand the size of the game tree abstraction in certain subtrees. We finally work on online learning techniques to allow a bot to adapt to multiple opponents by varying the aggressiveness of bot play. We find that both strategy stitching and online learning are effective in increasing bot performance.