Mastering Reconnaissance Blind Chess with Reinforcement Learning

Undergraduate Research Thesis at Georgia Tech

September 2018 - May 2020
Abstract

Research within Artificial Intelligence has often set goals of being able to autonomously play games (e.g., Chess or Go) at or above human level. Novel machine learning-based agents have recently made advances in the state-of-the-art by achieving superhuman performance in increasingly complicated games. We believe that solving imperfect information games (i.e., games where you do not have full knowledge of the opponent's activities) should be the next goal in Artificial Intelligence research. We study Reconnaissance Blind Multi-Chess (RBMC), an imperfect information variant of Chess, which comes with a novel set of challenges that must be overcome before a computer can attain superhuman performance. Prior works have largely focused on reducing the problem to a game of standard Chess (i.e., with perfect information) by attempting to determine the true state of the Chessboard. This procedure separates the problem of acquiring and applying gathered information from the move policy, allowing existing Chess agents to be used to choose nearly optimal moves. In contrast, our method trains a triple-headed neural network through self-play reinforcement learning, handling the information-gathering process, and move process within one model. Since this agent does not attempt to solve a restricted version of the problem, the algorithm is able to execute strategies based on the imperfect information aspect of the game. We believe that such a learning method, given enough training time, should be able to outperform agents that simply reduce the problem to a standard game of Chess. In this thesis, we explore this hypothesis and algorithms for playing RBMC.

Keywords: Reconnaissance Blind Multi-Chess, Reinforcement Learning.