Author: Robert Hyatt
Date: 11:11:53 06/01/04
Go up one level in this thread
On June 01, 2004 at 11:04:52, Sune Fischer wrote: >On June 01, 2004 at 10:07:15, José Carlos wrote: > > >> Sure it is. But the same could be said about time management. That module can >>be developed/investigated apart from the rest of the program. Or search and >>eval. Could work together (some programs prune based on eval) but a simple >>interface is enough and search and eval can be researched as different modules. >> Note that my point is that book related tools can be also subject of a most >>interesting research. Learning, for example. You have a limited space (you don't >>want your learned data to get huge) and some fuzzy information (this line looks >>promising or bad). > >It's not learning in an intelligent way, if you displace just one piece one >square it will be a completely different line to the book. Here, who cares? It _is_ learning in a very restrictive case. IE scientists teach mice to run thru mazes _exactly_ the same way, and you won't find anyone saying they are not "learning". > >Book learning is really more of a database management, learning is just too >fancy a word for something so relative primitive. Does it matter? IE is there some threshold we have to cross and then you say "ok, book learning is now a vital part of the chessplaying system and can't be turned off."??? What _is_ that threshold. Why do you infer that mice can't learn to navigate a complex maze, since they learn the _same_ way. They don't study the maze from above and discover a path that they use once they enter the maze. They just navigate it by trial and error and learn how to get through. That is _exactly_ how book learning works. yes it will probably get more sophisticated in the future. IE I already use search results to tune book probabilities, as well as using the more crude win/lose result to mark lines as "don't play". IE what I am doing is above what the mice do, and it will get better over time, too... > >An intelligent learner would adjust the game play itself, ie. parameter weights > would be tuned so that entire classes of openings were avoided in the future. I do this already... in a different way... >This could also be applied throughout the whole game and not just the opening. as in "position learning" as a first crude step? > >But in a small tournament of just 10-15 games you can't really improve the book >a whole lot, improving the book is really something you have to do _prior_ to >the tournament. >The only hope once the tournament begins is to try and repeat games, and that's >just lame and annoying. Duplicate games is a waste of everyones time and >patience, they should not be allowed to count twice anyway, one game - one win. > Wrong. _DEAD_ wrong. Against the computer I would probably play the most popular opening move every time. If it walks into a trap, I will lose 1/2 of the games since you have cleverly disabled my defense against this. Even though a Human would not play the same moves every time, the computer is forced to do so because of some twisted idea that "book learning is not part of the engine." Strange, but the first time I lost an evans gambit as black, I chose to not play that again for a while. I didn't study anything. I didn't think about it. I didn't ask anyone, I just didn't play that opening as black. It seems that I violates some golden rule by not doing so, according to this thread... > >> You can use information about your opponent (rating reported >>by winboard, or name of well known opponents). You make your decisions upon >>statistic information (a games database + your own games), the result of your >>search (this position looks good but I've lost the game), the game (I think I >>made a mistake later but this position is acceptable), your opponent's moves >>(his first move out of my book just killed me, I'll add to my own book)... >> There's a huge universe to research about book, and it is interesting if >>you're ready to think carefully about it. > >You argue well but it's not exactly rocket science, not compared to some of the >other challenges in chess programming. ;) There are many challenges. And there is nothing wrong with taking baby-steps toward solving more than one at the same time... > >> And finally, competition is about winning games under the rules. Kasparov can >>repeat the same opening against a program with no learning, and kill it 200-0 >>with only two different games. That program looks _stupid_ to the world. If you >>change Kasparov in that example for another program, you got a smart program and >>a stupid program. >> But as I told to Sune, it's a matter of taste. I like the program to do >>everything but moving the wood pieces on the board! > >There's gotta be a better and more interesting way than to just repeat won >games, if that's all there is to it then I'll happily swallow my duplicate >losses with pride. You are lost in the wrong world. It is not nearly so much about repeating "won" games as it is about not repeating _lost_ games. Perhaps that is the point you are missing. I would happily give up "won learning". But not "lost learning". >;) > >-S. >> José C.
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.