This thesis focuses on using neural networks to extract useful features from a poker game. These features will be used to exploit opponent behavior in playing poker. The opponent modeling network was built by Xun Li in prior work and this thesis is about exploring a different game state representation in the network. Currently, the input features are hand-made using common poker knowledge. The research is to let the computer figure out which features are important from the raw state of the poker game instead of using the hand-made features. Just as neural networks are proven to work better for extracting visual features for various applications, the project’s goal is to prove that for poker game representation. This research focuses on state representation only and actual gameplay (with the new representation) needs to be explored in future work. The model built in this thesis was able to represent most of the poker game state (except bet sizes) by significantly compressing it (through feature learning) while not losing much information. Moreover, the compressed representation is found to be similar yet better than the handcrafted input features because it packs more nuanced information than the latter. In the future, the model can be extended to include bet sizes and tested out in actual poker gameplay.

PDF

Technical Report, Department of Computer Sciences, The University of Texas at Austin.

@techreport{nagineni:thesis2018, title={Learning Useful Features For Poker}, author={Arjun Nagineni}, address={Austin, TX}, institution={Department of Computer Sciences, The University of Texas at Austin}, type={Undergraduate Honors Thesis}, url="http://www.cs.utexas.edu/users/ai-lab?nagineni:thesis18", year={2018} }