MCSIG GamSimTechMeth

Monte Carlo Search in Games

Project Leader Project Co-leader
Jonathan Schaeffer
University of Alberta
Holger Hoos
University of British Columbia

Fundamental improvements in Monte Carlo Tree Search algorithms may generalize from games like Go to new application domains where achieving intelligent behavior for artificial characters and opponents is key. Virtual worlds have clear rules and boundaries, with controllable complexity, and offer a stepping stone to less well-defined real-world applications. 

Automatic divide-and-conquer methods to avoid global search for large problems, computer-aided algorithm design, better incorporation of domain knowledge, and improved design principles are being investigated. “Imperfect information” games such as Hearts are expected to provide particular insights. Using multiple cores to increase decision quality and evaluate game positions plays a key role. 

MCSIG will improve decision making for imperfect information games and in the presence of uncertainty, infer hidden state from move sequences, and model opponents’ weaknesses.