Author: Robert Hyatt
Date: 05:25:44 06/22/00
Go up one level in this thread
On June 22, 2000 at 00:42:31, pavel wrote: >Hi guys, > I have been testing with crafty for some time now, and have come to the >position where i need experienced answer for my questions. I was wondering how >much elo differance will crafty have with hash table, positional.lrn , pondering >and without these? I mean in each case eg, hash size will crafty play better >game with 30mb hash than 10 mb hash? In the middlegame, crafty's search speed will vary by a factor of roughly 2X. If the hash is too small, the speed will be about 1/2 the speed if the hash table is large enough. > will crafty play better game with >positional learning than no positional learning at all? It is simply harder to get it to play the same game over and over, even if the book is very narrow (or non-existant). For a single game, this has absolutely no effect. Its primary focus is to prevent repeated losses when someone finds a way to take it out of book quickly, where the program would then tend to play the same moves over and over beyond that point. > same with pondering. It will probably predict correctly about 50% of the time. Which means it will save time in 50% of the positions, at most. This is a significant performance improvement. >In >each case how much (estimated) elo differance will occur? in the first and last case, either is worth 50-70 Elo rating points, roughly. For #2 it depends. If you try to play the same opening, over and over, with no (or limited) book, then this can be worth hundreds of rating points since it won't repeat lost games. > Currently I am testing crafty learning file by making it analyze 100s of >games played by GM (downloaded 1000s of games already by anand and kasparov). >The analyzing are done under CDB database program, (note under CDB if analyzed >with crafty it stores learned position in both book.lrn and positional.lrn.) as >a result the learn file has increased to something around 238kb (impressive? >....nah!!). I was wondering if its worth the trouble and as a result make crafty >play "better"?!. probably not worth anything at all. It is a defense against someone attacking the deterministic behavior of a chess program. It isn't intended to learn by going over games played by _others_. >thanks for your comments >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.