A Reinforcement Learning Approach For UH Leduc Poker

Poker, especially Texas Hold’em Poker, is a challenging game and top professionals win large amounts of money at international Poker tournaments. Consequently, Poker has been a focus of AI research to develop agents that play Poker intelligently. Challenges of Poker include partial observability, the need for probabilistic reasoning as Poker hands are dealt randomly, the difficulty to deal with an unknown adversary, the capability to bluff, and the difficulty of assessing the quality of a Poker hand in a particular game context. Leduc Hold’em Poker is a popular, much simpler variant of Texas Hold’em Poker and is used a lot in academic research. This work centers on UH Leduc Poker, a slightly more complicated variant of Leduc Hold’em Poker. The goal of this thesis work is the design, implementation, and evaluation of an intelligent agent for UH Leduc Poker, relying on a reinforcement learning approach. In particular, our approach employs Deep Q-Learning and the agent is implemented by using TensorFlow in Python. The UH Leduc Poker Agent is trained by playing tournaments against a fixed policy agent that plays smartly, according to the quality of its hand and the current state of the game. We also investigate the influence of different reinforcement learning parameters on the agent's performance. Finally, we conducted experiments that assess how well the UH Leduc Poker agent plays against some fixed policy agents as well as human beings.