Engineering Commons^™

A Deep Learning Agent For Games With Hidden Information, Robert A. Mills

Senior Projects Spring 2018

The goal of this project is to develop an agent capable of playing a particular game at an above average human level. In order to do so we investigated reinforcement and deep learning techniques for making decisions in discrete action spaces with hidden information. The methods we used to accomplish this goal include a standard word2vec implementation, an alpha-beta minimax tree search, and an LSTM network to evaluate game states. Given just the rules of the game and a vector representation of the game states, the agent learned to play the game by competitive self play. The emergent behavior from …