Learning Patterns in the Game of Go

This thesis introduces concepts for learning expert human positional classifiers in the game of Go. While most studies in the field of Go focus on learning methods to find the best move, this thesis focuses on learning methods for evaluating the quality of Go stones in a position. A data set of 2,638 Go Tesuji problems has been created. For these problems expert human classifiers have been recorded using a first order logic annotation extension developed for the current standardized format SGF (Smart Go File). The learning of the human classifiers has been conducted with Support Vector Machines using low level features such as the Relative Subgraph Features (RSF) from the Common Fate Graph (CFG). Relative Subgraph Path Features (RSPF) have been developed for learning connectivity and proved to be an appropriate representation for the learning task.