Institute of Information Science, Academia Sinica

Computer games have long been an important testbed for AI research. AlphaGo's success further demonstrated the potential of search-based reinforcement learning (RL) and motivated the development of more general zero-knowledge RL methods. This talk presents research on planning with zero-knowledge RL in games, focusing on AlphaZero- and MuZero-based methods. It first reviews the foundations, including the RL framework, Monte Carlo tree search (MCTS), policy and value networks, self-play and optimization, and dynamics learning. It then highlights selected research on effective, efficient, and trustworthy zero-knowledge RL methods, including strength adjustment for MCTS, Gumbel Stochastic MuZero, the MiniZero framework, AlphaZero-based game solving, interpretability of MuZero's learned model, and OptionZero for efficient planning with learned options.

Dr. Hung Guei is a Postdoctoral Scholar at the Institute of Information Science, Academia Sinica, Taiwan. He received his Ph.D. in Computer Science from National Yang Ming Chiao Tung University in 2023. His research interests include reinforcement learning, computer games, and artificial intelligence. His previous research on zero-knowledge reinforcement learning for computer games has yielded several significant outcomes, including publications in IEEE Transactions, such as IEEE CIM, IEEE TAI, and IEEE T-G; papers at top-tier AI conferences like AAAI, ICLR, and NeurIPS; and game-playing programs that have won international competitions.