訊息公告

【數據科學系列演講】2022.11.16 (三) 15:30-17:20@EC115/講者:魏廷翰博士/講題:Title:Exact Solutions in the Age of Deep Learning

 

講題: Exact Solutions in the Age of Deep Learning

主講人:魏廷翰 (Ting-Han Wei)博士

Postdoctoral Fellow in the Department of Computing Science at University of Alberta

時間:2022.11.16 (Wed.) 15:30-17:20

地點:工程三館 EC115

主持人:曾新穆教授

 

Abstract

Search is one of the key tools in the artificial intelligence toolbox. Being able to navigate large search spaces intelligently can play a major role in solving difficult problems. A traditional approach combines fast heuristics with efficient search algorithms to produce exact solutions, e.g. Boolean satisfiability problems, game solving, etc.

On the other hand, relatively slower but more accurate machine learned heuristics have brought about major milestones, the best example of which is the AlphaZero family of algorithms. More specifically, AlphaZero combines Monte Carlo tree search with deep reinforcement learning to achieve superhuman level strength in heuristic play for a wide range of games.

At the University of Alberta, under the supervision of Prof. Martin Müller, we explore many research topics surrounding exact and machine-learned methods. In this talk, I will introduce a number of our projects, sorted into three categories:

  1. Integrating exact solution methods with AlphaZero-like architectures
  2. Evaluating deep reinforcement learning against exact solutions
  3. Heuristic search and learning in imperfect domains

 

Speaker Bio

Ting-Han Wei is currently a postdoctoral fellow in the Department of Computing Science at University of Alberta. He received his Ph.D. in Computer Science from National Chiao Tung University in 2019, where he worked on analyzing the game of Connect6 using distributed computing. His research interests include artificial intelligence, computer games, and grid computing. He also published several papers in top-tier conferences, such as AAAI, IJCAI and ICLR. He is currently working with Prof. Martin Müller to develop a strong small board Go solver to determine the theoretical outcome of 6x6 Go.