Author: Robert Hyatt
Date: 14:11:46 06/03/98
Go up one level in this thread
On June 03, 1998 at 12:52:29, Komputer Korner wrote: >On June 03, 1998 at 08:42:37, Robert Hyatt wrote: > >>On June 03, 1998 at 00:51:13, Komputer Korner wrote: >> >>> >>>Okay I will use the KK Kup book from now on, but when I used the small >>>book that was created from a 400 game Morphy pgn file from M-Chess 7.1, >>>I used the n parameter as well. >> >> >>that's not really a "book". Because 99% of the moves will not be >>trusted since they weren't played more than once, and it also won't >>give much "coverage" for the normal openings either.. > >Doesn't the "n" parameter mean that any n > 1 means that all moves must >have been played at least n times. If so then Crafty will have enough >observations to give each move a rating, so I don't understand your yes. But to choose between the different moves, crafty uses the frequency they were played, their win/lose ratio, their static evaluation, and so forth, with frequency and win/lose being major terms. If there are only three moves, each with a count of 2 or 3, exactly what is it going to use to choose which of those moves leads to a reasonable position? If you make a book from 100,000 games, or 330,000 games like the KKUP book, it still ends up being very small, but it has an idea of how often e4, d4, and so forth are played, generally. Without that guidence, it is basically choosing randomly from the possible opening moves you supply. And many of those random choices will be gross mistakes... >comments about this. >I have seen the following happen more than once. I built the book with >the n parameter at least = 3. "WCrafty XX uses its book for a certain x >number of games and then after that, refuses to use its book anymore. So >I rebuilt it using another book/set of games. However, it may have been >because it decided that all lines were bad after losing to Comet. So I >will try the KK Kup book from now on. Human chess masters add to their >books and play from their books even though there is a possibility of >the line becoming refuted through deeper analysis. Speed increases on yes they do... but computers can't read informant and make decisions based on the comments it contains. All we can do is poke hundreds of thousands of games into it, and let it use information from each game to aid it in deciding whether to try that opening or not. And then learning will either reinforce or overrule that decision. But if you are playing matches with comet with a book so small that it can learn that everything is bad, your book is worthless... and it would be interesting to know how many unique positions it actually contains... >the order of doubling every 1.5-2 years will not enable many refutations >of book lines that have been played successfully by Crafty. If it does >happen that Crafty loses a game from one of these lines then Crafty can >always back the analysis up to see where the losing move was and adjust >its book accordingly if the book was at fault. Crafty would have lost >that game anyway assuming it's knowledge and search were the same, >because the line would only have been added to Crafty's book (if you >will implement this feature)because and THIS IS IMPORTANT, CRAFTY WOULD >HAVE calculated the line anyway in a previous game. So it might as well >be added to Crafty's book until it is refuted. This will save thinking >time. Of course, any time the search algorithm or knowledge is changed >then a new book should be created that will optimize the play of the NEW >Crafty. BUT assuming a constant Crafty it would be very interesting to >see Crafty create a book of its own (without any limits) based on its >own play. your base assumption is invalid. "constant crafty" is (at present) an oxymoron, which is why I don't add to the book (not to mention the difficulty of doing so with the present *very fast* format) and that is also why I delete the position.* files regularly as well... > This would be like an AI project but an AI project that would >be strictly a chess opening book project.
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.