Poker with Reinforcement Learning Agents

Overview

I created a simplified Texas Hold'em environmnet for training reinforcement learning agents using Deep Counterfactual Regret Minimization (CFR). Additionally, nn-holdem and rlcard were enormously helpful for this project.

Poker and Machine Learning

Applications in game theory

• Poker is a type of imperfect information game providing an opportunity to use machine learning agents to find winning strategies.
• Created a simplified Texas Hold'em environment to train a CFR agent.
• The CFR agent showed improved performance over fixed strategy agents (i.e. no learning involved) to win a majority of hands.
• The rlcard repo allowed me to upload my agent and have it play against other AIs.
• Further development:

-Train the model on real-world poker hand data using LSTM to analyze player trends in series of hands.
-Use image recognition to identify a player's "tells" when they are bluffing.
-Continue to develop a full Texas Hold-em game environment.

* Reinforcement learning could be applied to any real-world imperfect information situation, like international business or trade negotiations. Rewards and expected responses could be modeled to determine strategies that maximize rewards.

Resources:

• rlcard (https://github.com/datamllab/rlcard) @article{zha2019rlcard, title={RLCard: A Toolkit for Reinforcement Learning in Card Games}, author={Zha, Daochen and Lai, Kwei-Herng and Cao, Yuanpu and Huang, Songyi and Wei, Ruzhe and Guo, Junyu and Hu, Xia}, journal={arXiv preprint arXiv:1910.04376}, year={2019} }
• This article was extremely helpful in constructing my own agent: https://medium.com/emergent-future/simple-reinforcement-learning-with-tensorflow-part-1-5-contextual-bandits-bff01d1aad9c
• environ setup: https://github.com/alexbeloi/nn-holdem