Using Reinforcement Learning in Chess Engines

Up until recently, the use of reinforcement learning (RL) in chess programming has been problematic and failed to yield the expected results. The breakthrough was finally achieved through Gerald’s Tesauros work on backgammon, which resulted in a program that could beat the world champion of backgammon in the majority of the matches they played. Our chess engine proved that reinforcement learning in combination with the classification of board state leads to a notable improvement, when compared with other engines that only use reinforcement learning, such as KnightCap. We extended KnightCap’s learning algorithm by using a bigger and more complete board state database, and adjusting and optimizing the coefficients for each position class individually. A clear enhancement of our engine’s learning and playing skills is reached after only a few trained games.