Playing Good-Quality Games with Weak Players by Combining Programs with Different Roles
Chu-Hsuan Hsueh, Kokolo Ikeda · 2022
Computer programs have become stronger than top-rated human players in several games. However, weak players may not enjoy playing against these strong programs. In this study, we propose combining two programs with different roles to create programs suitable for weak players. We use a superhuman program that generates candidate moves and evaluates how good the moves are, as well as a program that evaluates the moves’ naturalness. We implement an instance for Go, which employs a superhuman program, KataGo, and a neural network trained using human games. Experiments show that the proposed method is promising for playing good-quality games with weak players.