Move Prediction in Go – Modelling Feature Interactions Using Latent Factors

Abstract

Abstract. Move prediction systems have always been part of strong Go programs. Recent research has revealed that taking interactions between features into account improves the performance of move predictions. In this paper, a factorization model is applied and a supervised learning algorithm, Latent Factor Ranking (LFR), which enables to consider these interactions, is introduced. Its superiority will be demonstrated in comparison to other state-of-the-art Go move predictors. LFR improves accuracy by 3% over current state-of-the-art Go move predictors on average and by 5% in the middle-and endgame of a game. Depending on the dimensionality of the shared, latent factor vector, an overall accuracy of over 41% is achieved.

Common fate graph patterns in Monte Carlo Tree Search for computer go

“…The CFG patterns increase the prediction accuracy of the tree-search-patterns to 42.1 % which is even larger than the results published in [18] who use latent factors to model interactions between features and whose results are among the best published so far. Figure 7 the log-likelihood of selecting the expert move is plotted.…”

Convolutional Monte Carlo Rollouts in Go

“…Coulom (2007a) computed the likelihood of patterns and local features with a Bradley-Terry model. Wistuba & Schmidt-Thieme (2013) used latent factor ranking to achieve move prediction accuracy of 41%.…”

Factorization Ranking Model for Move Prediction in the Game of Go

“…Most state-of-the-art computer Go programs rely heavily on such knowledge (Browne et al 2012). Using simple features such as patterns, popular offline training algorithms such as Minorization-Maximization (MM) ) and Latent Factor Ranking (LFR) (Wistuba and Schmidt-Thieme 2013) produce models that are fast enough to evaluate at every node in a MCTS game tree.…”

“…The current paper proposes and evaluates the Factorization Bradley-Terry model for move prediction. In popular move prediction algorithms for the game of Go, such as Wistuba and Schmidt-Thieme 2013), each move in a game is modeled as a group of features which describe it. Best move selection then becomes a competition among the groups representing all legal moves in a Go position.…”

“…The major innovation of the proposed Factorization Bradley-Terry (FBT) model is to combine the strengths of the two leading approaches Wistuba and Schmidt-Thieme 2013), and consider the interaction between individuals within the same group as part of a probability-based framework, which is inspired by the Factorization Machine model (Rendle 2010).…”