Author: Robert Hyatt
Date: 07:51:04 01/10/99
Go up one level in this thread
On January 09, 1999 at 17:58:11, James T. Walker wrote: >On January 09, 1999 at 06:02:21, blass uri wrote: > >> >>On January 09, 1999 at 05:40:14, Thom Perry wrote: >> >>>On January 08, 1999 at 11:46:01, blass uri wrote: >>> >>>> >>>>I checked tiger's results and I found that the results were: >>>> >>>>8.5 out of 18 (first 2 games against every player with white and black) >>>>10 out of 18(games 3-4 against the same 9 opponents) >>>>10.5 out of 18(games 5-6 against the same 9 opponents) >>>>11.5 out of 18 (games 7-8 against the same 9 opponents) >>>>10 out of 18(the last 2 games with different colours against the same players) >>>> >>>>I do not see that tiger is getting less points when there are more games. >>>> >>>>Fritz532 tried to repeat with black an opening that it won but tiger refused to >>>>repeat the same game(It is probably not deterministic) >>>> >>>>Fritz532 won the first 2 games of the same opening with black but in the third >>>>game when fritz repeated the same opening tiger drew. >>>> >>>>Tiger won the last game against Fritz532 with white(Fritz did not try to use the >>>>same opening that tiger drew against it) >>>> >>>>Uri >>> >>>Looks to me like Tiger is learning. Like fine wine, it is getting better >>>with age. >> >>Christophe(The programmer of chesstiger) said that it has not a learning >>feature. >> >>Uri >>>>>>>>>>>>>. >I'm still not convinced of the usefullness of book learning. I don't know how >many programs have "hash learning" like Nimzo99. But with book learning, I >haven't seen an explanation of how it works so I can't put it down too much. >The thing is, with todays programs which have hundreds of thousands of book >moves which make up several thousand variations, they can play a different >variation everyday for the next ten years and not repeat. In that case what >good does it do to learn not to use a particular variation again? I guess I'll >have to wait for Dr. Hyatt's paper to ICCA. >Jim Walker I'll give this a quick shot. This paper should be in the March issue if I understood Jaap correctly. It was being rushed to completion for the December issue, but they ended up with too many articles. Which gives me time to do a bit more polishing for March. There are two types of book learning in crafty. The simplest is 'result learning' and simply says "if a book line leads to a loss, then, working backward from the end of the book line, find the last point where there was a choice to be made, and mark the choice actually played as 'never play'." The more complex form simply uses the first 10 searches out of book to come up with a 'learned value' that is based on the search scores, the opponent rating, and the search depth reached. This learn value is propogated back up the tree, so that if the learned value is negative, crafty will tend to avoid those moves (depending on how 'strongly' you have learning weights set). It does work, and works well. My book is limited to moves that were played at least 5 times each, and for every 2 losses there must have been at least 1 win, otherwise the book move is discarded. And this *still* leaves bad moves in the book (I will see several -2.5 to -3.5 scores right out of book.) Now, I ignore those as Crafty marks them as bad automatically... And if you want, when it learns an opening is bad for black, it will try to play it as white also.
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.