Bayesian pattern ranking for move prediction in the game of Go

Abstract

We investigate the problem of learning to predict moves in the board game of Go from game records of expert players. In particular, we obtain a probability distribution over legal moves for professional play in a given position. This distribution has numerous applications in computer Go, including serving as an efficient stand-alone Go player. It would also be effective as a move selector and move sorter for game tree search and as a training tool for Go players. Our method has two major components: a) a pattern extraction scheme for efficiently harvesting patterns of given size and shape from expert game records and b) a Bayesian learning algorithm (in two variants) that learns a distribution over the values of a move given a board position based on the local pattern context. The system is trained on 181,000 expert games and shows excellent prediction performance as indicated by its ability to perfectly predict the moves made by professional Go players in 34% of test positions.

Move Prediction in Go – Modelling Feature Interactions Using Latent Factors

“…Most move predictors for Go are either using Neural Networks [6,7] or are estimating ratings for moves using the Bradley Terry (BT) model or related models [3,4,8]. Latter mentioned approaches model each move decision as a competition between players, the move chosen by the human expert player is then the winning player and its value is updated accordingly.…”

“…Another possibility to divide the Go move predictors into two classes is how they consider interactions between features. There are two variants, one models the full-interaction of all features [3,9] and the others do not consider them at all [4,6,8].…”

“…In practice, approaches not considering feature interactions at all proved to be more accurate. Stern's [3] full-interaction model used a learning set of 181,000 games with 4 feature groups but only predicted 34% of the moves correctly. Using the same approach with no interaction, Wistuba et al [4] has shown that easily 9 feature groups can be used and, using a learning set of only 10,000 games, 37% of moves were predicted correctly.…”

“…Shape features take the shape around a specific intersection on the board into account, non-shape features are every other kind of features in a move situation you can imagine. How shapes are extracted, harvested and represented is explained very well in [3].…”

“…Upfront, the strength of each feature is learned with various supervised learning algorithms. The prediction can be improved by using additional features, but as seen in [3,4] it can also be improved by considering the impact of feature interactions.…”