Previous: Expected Outcome
Up: Discussion
Next: Evaluation Function Learning
Previous Page: Expected Outcome
Next Page: Evaluation Function Learning

Automatic Feature Generation

There has recently been much progress in developing programs which generate features automatically from the rules of games [\protect\citenameFawcett and Utgoff, 1992][\protect\citenameCallan and Utgoff, 1991][\protect\citenamede Grey, 1985]. When applied to chess such programs produce features which count the number of chess pieces of each type, and when applied to Othello they produce features which measure different aspects of positions which are correlated with mobility. The methods operate on any problems encoded in an extended logical representation, and are more general than the methods currently used by METAGAMER. However, these methods do not generate the values of these features, and instead serve as input to systems which may learn their weights from experience or through observation of expert problem-solving. While METAGAMER's analysis is specialised to the class of symmetric chess-like games, and thus less general than these other methods, it produces piece values which are immediately useful, even for a program which does not perform any learning.