Understanding the Success of RL-based Chess Agents

Inspired by the success of AlphaZero, my research investigates RL in board games, particularly focusing on the contrasting algorithms of two AI chess agents: Stockfish and Leela Chess Zero (Lc0). The study examines the strengths and potential vulnerabilities of these algorithms, with a special interest in how they perform under various conditions. A key finding is Lc0’s consistent performance over Stockfish, leading to the development of a novel agent. This new agent combines the intuitive aspects of DNNs with the strategic depth of minimax search, showing promising results in preliminary tests. This finding hints at exciting possibilities for blending self-play in RL with classic search techniques in AI.

Please see below as the slides I used for a presentation at Davidson College.