This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Although AlphaZero has achieved superhuman performance in board games, it struggles in scenarios that require a comprehensive understanding of the entire board (e.g., long-term pattern recognition in Go). In this paper, we propose ResTNet, which alternates residual blocks and Transformer blocks to bridge local and global knowledge. ResTNet improves the winning rate in several board games (9x9 Go: 54.6% → 60.8%, 19x19 Go: 53.6% → 60.9%, 19x19 hex: 50.4% → 58.0%) and effectively handles long-term patterns such as circle and ladder patterns in 19x19 Go. It reduces the mean squared error of circle pattern recognition from 2.58 to 1.07, reduces the attack probability against adversarial programs from 70.44% to 23.91%, and improves the accuracy of ladder pattern recognition from 59.15% to 80.01%. Visualization of attention maps provides insight into AlphaZero's decision-making process by capturing important game concepts in both Go and Hex. ResTNet presents a promising approach to local and global knowledge integration, paving the way for more effective AlphaZero-based algorithms for board games. The code can be found at https://rlg.iis.sinica.edu.tw/papers/restnet .
