Download PDF Version

    Abstract: With the success of Monte Carlo Tree Search, the game of Go has become a focus of games research. Recently, deep convolutional neural networks have achieved human-level performance in predicting master moves. Even before that, machine learning techniques have been used very successfully as an automated way to improve the domain knowledge in Go programs. Go programs have now reached a level close to top amateur players. In order to challenge professional level players, we must combine the three pillars of modern Go programs - search, knowledge, and simulation - in a high performance system, possibly running on massively parallel hardware. This talk will summarize recent progress in this exciting field, and outline a research strategy for boosting the performance of Go programs to the next level. 

    Martin Müller is a Professor and the Associate Chair (Research) in the Department of Computing Science at the University of Alberta in Edmonton, Canada. His research covers several areas of heuristic search, including game tree search, domain-independent planning, and combinatorial games. His recent main focus is on understanding exploration in heuristic search. He has worked on computer Go for thirty years, and leads the development of the open source games software framework Fuego. In 2009, Fuego was the first program to beat a top level professional in an even game on the 9x9 board. Games programs based on the Fuego framework have won many major international competitions, including the Computer Olympiad and the UEC Cup. In work in the field of AI planning, Müller’s research group has developed a series of internationally successful planning systems based on macro learning and on Monte Carlo random walks. He has numerous publications in top quality venues and major conferences in the fields of computer games, planning and general AI. He heads the computer Go and planning groups, which are part of the world class AI research group at the University of Alberta.