Machine-Learning of Shape Names for the Game of Go

Abstract

Computer Go programs with only a 4-stone handicap have recently defeated professional humans. Now that the strength of Go programs is sufficiently close to that of humans, a new target in artificial intelligence is to develop programs able to provide commentary on Go games. A fundamental difficulty in this development is to learn the terminology of Go, which is often not well defined. An example is the problem of naming shapes such as Atari, Attachment or Hane. In this research, our goal is to allow a program to label relevant moves with an associated shape name. We use machine learning to deduce these names based on local patterns of stones. First, strong amateur players recorded for each game move the associated shape name, using a pre-selected list of 71 terms. Next, these records were used to train a supervised machine learning algorithm. The result is a program able to output the shape name from the local patterns of stones. Including other Go features such as change in liberties improved the performance. Humans agreed on a shape name with a rate of about 82 %. Our algorithm achieved a similar performance, picking the name most preferred by the humans with a rate of about 82 %. This performance is a first step towards a program that is able to communicate with human players in a game review or match.