Author: Miguel A. Ballicora
Date: 11:31:22 01/18/01
In my program, I have not added a book yet. So, I decided to include a simple learning technique so it will not repeat the same losing game over and over. Every time it plays a move, it compares the evaluation returned by the search with the evaluation returned the previous move it played. It the current evaluation is lower by 0.20 paws or more, there is an indication that there is something that the program did not see before. So, I store the position with a move in a hash table (1Mb) reserved only for this purpose. At the end of the sesion (or game) the program store this table as a binary file that retrieves every time it starts a game or series of games. When the program searches, it probes for transposition in the regular hash table. If it does not find anything, it looks for the position in the 1Mb hash table used for "learning". In this way, it remembers not so good positions that the program fell in previous games and try to avoid them. I did not test this much, but it seems that it adds a little of a good random play to my program even without an opening book. Even after winning a game changes the play because it always find a suboptimal move from the previous game. I am sure that this has been tried before (or something similar). Has been tried in this context? what do you think? I think that it is a good idea if no book is present. So far, I like it. I tried another one that I did not like much so far (did not test it thoroughly, though): In the evaluation function, I added a random term that depended on the hash signature of that position. In that way, it was random but *reproducible*. (the seed, is determined at the beginning of the game from the time). However, the play was not very random or started to play crazy. Maybe the random numbers used were not good or I did not tuned properly the parameters. More to test. Miguel
This page took 0.02 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.