TIGP (SNHCC) -- AI in Computer Games: Game Solving, Strength Modeling, and Interpretability
- 講者施仲晉 教授 (國立中正大學資訊工程學系)
邀請人：TIGP (SNHCC) - 時間2026-09-07 (Mon.) 14:00 ~ 16:00
- 地點Google Meet
線上串流
Google Meet 會議參加資訊
視訊通話連結：https://meet.google.com/tgp-rkxw-xvo
或撥打以下電話號碼：(US) +1 984-221-1697 PIN 碼：721 263 857#
更多電話號碼：https://tel.meet/tgp-rkxw-xvo?pin=5550471716548
視訊通話連結：https://meet.google.com/tgp-rkxw-xvo
或撥打以下電話號碼：(US) +1 984-221-1697 PIN 碼：721 263 857#
更多電話號碼：https://tel.meet/tgp-rkxw-xvo?pin=5550471716548
摘要
Beyond playing to win, game AI raises deeper questions: can we solve a game, model a player's skill, and explain why a move is good? This talk introduces three research directions that pursue exactly these goals.
First, in game solving, techniques such as Relevance-Zone search, the Proof Cost Network, and distributed solving frameworks make it possible to prove game outcomes efficiently and at scale, while guaranteeing correctness.
Second, in strength modeling, a learned Strength Estimator captures a player's skill level. This enables AI that adjusts its difficulty in a controllable, human-aligned way, rather than always playing as strongly as possible.
Third, in interpretability, a Residual–Transformer hybrid model combines local patterns with global board structure to produce interpretable signals about positional factors and move quality, helping us understand why a move is good.
Together, these directions point toward game-AI systems that are accurate, adaptive, and interpretable.
First, in game solving, techniques such as Relevance-Zone search, the Proof Cost Network, and distributed solving frameworks make it possible to prove game outcomes efficiently and at scale, while guaranteeing correctness.
Second, in strength modeling, a learned Strength Estimator captures a player's skill level. This enables AI that adjusts its difficulty in a controllable, human-aligned way, rather than always playing as strongly as possible.
Third, in interpretability, a Residual–Transformer hybrid model combines local patterns with global board structure to produce interpretable signals about positional factors and move quality, helping us understand why a move is good.
Together, these directions point toward game-AI systems that are accurate, adaptive, and interpretable.