Abstract for HONS 02/20
Transfer Learning in Discrete Zero-Sum Games
Nathan Cleaver
Department of Computer Science and Software Engineering
University of Canterbury
Abstract
The construction of a general artificial intelligence is a long-standing goal of computer science. Transfer learning, the ability to transfer knowledge from one context to another, is an essential component of a general intelligence. This work investigates the effects of transfer learning in discrete, zero-sum games. We use distributed parallel Monte Carlo Tree Search to construct game-playing models for chess and a chess variant, Chessplus. Transfer learning is used to allow the chess model to play Chessplus, and results measured. To this end, we also introduce a novel pruning algorithm for Monte Carlo Tree Search. We demonstrate that positive transfer learning in Monte Carlo Tree Search is possible, and in our work, highly successful. The model with transfer outperforms the model without, winning 47.4% of games and drawing 16.9%. Statistical analysis supports the validity of our results. Further analyses offer valuable insights into the specific benefits of transfer between these two domains.
