Author: Robert Hyatt
Date: 11:28:38 07/27/00
Go up one level in this thread
On July 27, 2000 at 14:00:06, pavel wrote: >On July 27, 2000 at 13:41:24, Robert Hyatt wrote: > >>On July 27, 2000 at 13:10:44, pavel wrote: >> >>> >>>is it possible to make crafty acces the position.lrn the same way it accesses >>>the egtb? in this case, you dont have to worry about loading the lrn file every >>>time to the hash table, and also the size wont be a problem. since crafty can >>>deal with 20gb of egtb it shouldnt be a problem with few mbs of learning file. >>>Or am I wrong ? will it cost crafty some speed ? or it is not possible to code >>>is such way? >> >>The problem is this: And EGTB has all possible scores for a given set of >>pieces. The position.bin file has a few positions from wildly separated >>positions in many games. A hit is so unlikely it hardly ever happens. It >>is most useful when someone takes crafty out of book, and for the first 10 >>moves, Crafty is doing fine, but then they walk it into a trap it didn't >>see. Prior to "result learning" this could cause problems, as the book line >>looks fine, but it loses later. With position learning, it will _still_ >>play the same book line, but it will vary and play a different move somewhere >>before the point where it failed low and lost when the position was saved. >> >>In tournaments, I would not use it at all, since I would never let it play >>the same opening twice in one tournament... On ICC, it helps sometimes, but >>the book is wide enough that it isn't used very much at all. I doubt it would >>affect it much if it was turned off. > >I should have been more clear , I was interested in position.lrn not book.lrn. >so wont the concept be effective with position.lrn? I was talking about position.bin/position.lrn. Book learning only works in two cases: (1) after 10 moves out of book, the score is bad. It marks the latter part of the book line to avoid it next time; (2) after a game is lost, the latter part of the line is marked as bad. But in position learning, it is useful _after_ we are beyond the threshold where book learning works. It operates as I described by adding permanent entries to the hash table so that a repeat of this game will produce a different move. >then It wont deal with only the first 10 moves after book but throughout the >game? or there is a move limitation.? >I am so interested about crafty learning feature, I always facinated about >computer chess programs with *effective* learning features. and crafty is the >only one that has a "good" learning feature AFAIK. Also is there any other way >that learning can be tuned in to computer chess? which is much more effective >with fewer limitations? >also what is cooking with crafty? seen crafty17.13 play on chess.net >what are your future plans? like to share? >thanks for the explanations >pavel > >>>but if you look at the bright side, "if" crafty comes in the same position twice >>>things will be "faster for him to search", he will know the outcome of the game >>>within his evaluation. I know the possiblity of crafty coming to same position >>>is "dubious", but if you play 1000s of games doesnt the possibility gets >>>narrowed? also it is possible to make crafty "learn" through analysys within >>>CDB, its faster and effective IMO. or you can create a "universal" learning file >>>from thousands of games played by crafty online. >>>well I may be wrong since I a not an expert. >>>but IMO its not a bad idea............. yes IMO ;) >>>thanks >>>pavel. >>> >>>ps, ofcourse the post is mainly indicated for Dr, Hyaat, but others' comments >>>are appreciated. thanks. pavel
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.