Author: James Swafford
Date: 08:50:22 09/12/99
Go up one level in this thread
On September 12, 1999 at 01:45:28, James Robertson wrote: >My program currently book learns by storing the result of the game in each >appropriate book record (it increases the move's score if it won, and lowers it >if it loses). > >I just remembered someone mentioning something about how Crafty book learns >based only on the score it got _leaving_ the book (regardless of the result). Is >this true? Does it work better? > >Any help or comments on what your program does would be helpful. > >Thanks, >James I do that, but I also store a score. After the game is over, the engine researches the first X moves to a fixed depth, and stored the result *if* the move is in book. (If it isn't, the move isn't searched). The program won't play a move with < .25 pawn eval, unless it is stamped with an old version of the evaluator. There are a lot of improvements that could be made, though. For starters, I should append the moves that aren't in book to a "merge file," then periodically execute a merge to add them to the master book. Second, it is possible that during a game the engine will have already searched to depth X. In these cases, performing a research at the end of the game is stupid. In many cases, it could have had a score from a depth X+1 or greater if it used the search score from the game. I don't do anything with the win/loss results yet, but I do store them. -- James
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.