| Author |
Message |
Erik Zoltan (Zoltan)
| | Posted on Saturday, January 20, 2001 - 8:55 am: | |
Zillions is pretty darn smart about figuring out how much each piece is worth in a chess-like game. However the results could be improved if it learned from experience. Every game is different and the engine can't possibly come up with "perfect" piece weightings every time. Example: a light-squared bishop would be useless in checkers.
If there was a menu item called "Practice this game" then it could optimize the relative values of all its evaluation criteria. It would randomly make one or more changes in the "weights" of various evaluation criteria (e.g. how much a pawn is worth, or the bonus for attacking a royal piece, or whatever.) It would then play itself a couple of games at a very quick time limit, pitting the current weights against the modified weights. If the modified weights wins more games, then it might try a slower time limit, and if it still wins then it would adopt the modified weights, then repeat the process and continue learning. The user could leave it running for as long as they want, and the program would get better as a result. |
|