Author: Robert Hyatt
Date: 12:56:27 08/24/98
Go up one level in this thread
On August 23, 1998 at 19:45:20, Don Dailey wrote: >On August 21, 1998 at 23:55:07, blass uri wrote: > >>I think it can be a good idea to do a tournament of programs when >>every game between 2 computer programs will be replayed 3 times: >>1 time when both program have an opening book. >>1 time when only white has an opening book. >>and 1 time when only black has an opening book. >> >>By this tournament we can estimate the value of the books opening in elo. >> >>Uri > > >A reasonable idea. I have another idea for those situations where you >want to play long matches between just 2 machines to minimize the >effect of killer books. It's certainly not a new idea but a reasonable >one I think. You simply determine in advance how deep you can go in >the openings and then not allow duplicate games beyond this point. For >instance if you decide 8 moves provide adequate variety, then the >first time another game duplicates within the first 8 moves, (even if >the result is different) then you throw out the 2nd game and start over >until you get another unique opening. An opponent can still optimize >but this becomes much more difficult to do and the opponent program will >have to take more chances to do it right like having to put weaker moves >in his book to bait a specific opponent. > >Another similar idea is to decide in advance how many games should be >in the match (n) and play 2n games. Use the above idea but determine >what number of moves is just big enough to not exclude too many games. >This has the merit of not having to figure this out in advance but >requires much more testing overhead. > >- Don I have another alternative: How about we just continue to work on book-learning, which is still fairly infantile in the way things work, and let the programs fend for themselves? IE (and speaking only for myelsf here) my goal is to have a program that can stand by itself, against any sort of pranks you might throw at it... trying to repeat games... trying to use pre- calculated killer book lines... trying to coax it into dubious positions by offering a gambit pawn... It is too easy to find ways for a human to finagle a program so that it doesn't get burned. I find it *much* more interesting to finagle the program so that it finagles itself and self-avoids these silly messes. There are problems, because book-learning is not real good (at least my version) since a book can have millions of games in it, and you really ought to play each game once to confirm that it is playable or not... and that is *really* impossible... Someone suggested that the "crafty world" might do this by setting up each crafty clone to play only one specific opening and become a "expert" in it (since learned data is sharable with the import facility in Crafty). And this might be a solution... one only plays the French as black in response to 1. e4, and it would not take forever to play the various book lines and learn what is good and what is bad. But I like the "automatic" approach, rather than trying tournaments with no books, limited books, identical books, etc. Something tells me that Genius would "die" without the book it normally uses, because it has some lines that it handles very well, while other programs would probably not be happy. It doesn't make a lot of sense to spend time trying to test "parts" of a program, because it is really a complete entity, that includes the engine, the book, the endgame databases, learning information, and whatever else is part of the "thing". Also it seems more satisfying to solve something once, than to have to continually solve it over and over manually...
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.