TOP > 講演会 > 2009-01-27



場所:電気通信大学 西9号館3階AVホール

講演者:Remi Coulom氏(シャルルドゴール大学;CrazyStone作者)
講演1:Criticality: a Monte-Carlo Heuristic for Go Programs

When applying Monte-Carlo tree search in a Go-playing program, a variety of playout statistics can be used to guide the search. All-moves-as-first value, average score, or average point owner, in addition to winning frequency, provide useful information for the search algorithm. Criticality is another Monte-Carlo heuristic that estimates how important it is to own a point in order to win playouts. It was used succesfully as a pattern feature in Crazy Stone.

講演2:Local Quadratic Logistic Regression for Stochastic Optimization of Parameters

An important aspect of improving a game-playing program is the optimization of parameters from the observation of game outcomes. When the outcome of a game is binary (win/loss) or ternary (win/draw/loss), it provides a very noisy estimate of program strength. Estimating the probability of winning requires playing a very large number of games, and takes a lot of CPU time. Local Quadratic Logistic Regression is an efficient method to perform parameter optimization. It consists in fitting a quadratic response-surface model to the performance data. A statistical test of quadraticity is used to restrict this regression to a neighborhood of the optimal parameters. Extensive empirical testing with artificial functions demonstrates fast convergence. This method was used successfully to tune some parameters of Crazy Stone.