Strength Adjustment and Assessment for MCTS-Based Programs

• Main Authors: Liu, An-Jen, 劉安仁
• Other Authors: Wu, I-Chen
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/br8e7d

碩士 === 國立交通大學 === 資訊科學與工程研究所 === 107 === This paper proposes an approach to strength adjustment and assessment for MCTS-based game-playing programs. We modify an existing softmax policy with a strength index to choose moves. The most important modification is a mechanism which filters low quality moves by excluding those that have a lower simulation count than a pre-defined threshold ratio of the maximum simulation count. Through theoretical analysis, we show that the adjusted policy is guaranteed to choose moves exceeding a lower bound in strength by using a threshold ratio. The approach is applied to the Go programs ELF OpenGo and CGI. Experiment results show that the strength index is highly correlated to the empirical strength; namely, given a threshold ratio 0.1 as an example, the strength index is linearly related to the Elo rating with regression error 47.95 Elo. With an index value between $ m 2$, we can cover a strength range of about 800 Elo ratings. We then present three methods to adjust strength and assess opponents’ strengths dynamically. The strength adjustment and assessment methods were also tested in real-world scenarios with human players, ranging from professionals (strongest) to kyu rank amateurs (weakest). In addition to offering a wide range of strengths, our method can consistently play at the same strength given the same strength index, allowing players to choose appropriate opponents with ease. In regards to assessment, our method can automatically adjust its strength to its opponent's and in turn measure the opponent's strength accurately within 15 games. To our knowledge, this result is state-of-the-art in terms of the range of strengths in Elo rating while maintaining a controllable relationship between the strength and a strength index.