2019/02/26 by Céline Hocquette, Hocquette, Céline, Stephen Muggleton +2 · 1 voice
Computer Science · #Artificial Intelligence (cs.AI) #Artificial Intelligence in Games #Explainable Artificial Intelligence (XAI) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Reinforcement Learning in Robotics #cs.AI #cs.LG
paper · pdf · doi:10.48550/arxiv.1902.09835
openalex publication_date 2019/02/26 · arxiv published 2019/02/26 · arxiv updated 2019/02/26 · openalex created_date 2022/07/29 · openalex updated_date 2026/07/28
World-class human players have been outperformed in a number of complex two\nperson games (Go, Chess, Checkers) by Deep Reinforcement Learning systems.\nHowever, owing to tractability considerations minimax regret of a learning\nsystem cannot be evaluated in such games. In this paper we consider simple\ngames (Noughts-and-Crosses and Hexapawn) in which minimax regret can be\nefficiently evaluated. We use these games to compare Cumulative Minimax Regret\nfor variants of both standard and deep reinforcement learning against two\nvariants of a new Meta-Interpretive Learning system called MIGO. In our\nexperiments all tested variants of both normal and deep reinforcement learning\nhave worse performance (higher cumulative minimax regret) than both variants of\nMIGO on Noughts-and-Crosses and Hexapawn. Additionally, MIGO's learned rules\nare relatively easy to comprehend, and are demonstrated to achieve significant\ntransfer learning in both directions between Noughts-and-Crosses and Hexapawn.\n