Previous: Discussion
Up: Discussion
Next: Automatic Feature Generation
Previous Page: Discussion
Next Page: Automatic Feature Generation

Expected Outcome

Abramson [\protect\citenameAbramson, 1990] developed a technique for determining feature values based on predicting the expected-outcome of a position in which particular features (not only piece values) were present. The expected-outcome of a position is the fraction of games a player expects to win from a position if the rest of the game after that position were played randomly. He suggested that this method was an indirect means of measuring the mobility afforded by certain pieces. The method is statistical, computationally intensive, and requires playing out many thousands of games. On the other hand, the analysis performed by METAGAMER is a direct means of determining piece values, which follows from the application of general principles to the rules of a game. It took METAGAMER under one minute to derive piece values for each of the games discussed in this section, and it conducted the analysis without playing out even a single contest.