As the world's best chess players are computers since some years ago, computer scientists are now faced with the next challenge which is the ancient asian board game called go (chin.: wéiqí, kor.: baduk). Computing the best, or at least a good move to a given board position for the game of go is in the focus of researchers working on artificial intelligence since many years. While applying the same algorithms and ideas that were used for the chess game showed only little success, a monte carlo based approach brought about great improvements in the past few years. The best computer go programs have recently reached a good amateur playing strength.
In the game of go the number of possible moves a player can select from is very large compared to the playable moves in the chess game. Furthermore there are no good policies known to diminish the number of moves in a save way by classifying for example senseless or clearly bad moves. As a result of this we need to explore a very large game tree to compute a good follow up move to a given game situation. In a monte carlo based approach only single paths of the whole tree get explored and statistics about the outcomes are collected at tree nodes close to the root. We call this exploration of a single path from the root node to a leaf node of the game tree a monte carlo sample. The accuracy of the statistics and so the accuracy of the whole procedure increases with the number of samples done.
In the Gomputer project we aim to use special purpose hardware as FPGAs and GPGPUs to accelerate the monte carlo sample computation, thus exploit fine grained parallelism. On a higher level we intend to use many-core systems in a compute cluster to distribute the whole search. As increasing the number of monte carlo samples is only one side of the coin, investigations concerning the sample quality need to be made. Therefore the analysis and development of machine learning algorithms to extract game specific knowledge from professional game records that can be used during the sample generation thereafter becomes an important task for us.
Funding and Collaboration
The project is funded by Microsoft Research Cambridge by providing a Phd-Scholarship.
 "Modification of UCT with Patterns in Monte-Carlo Go", S. Gelly et al., Technical Report 6062, INRIA, 2006.