Author: Robert Hyatt
Date: 19:25:17 06/24/02
Go up one level in this thread
On June 24, 2002 at 11:37:08, William H Rogers wrote: >On June 24, 2002 at 10:43:34, Sven Reichard wrote: > >>On June 24, 2002 at 09:22:41, Vladimir Medvedev wrote: >snip... >>2) Of course the main structure of the evaluation function can't be changed >>(unless the program modifies its own source code and recompiles itself), but if >>the evaluation function is flexible enough it can discover new features and use >>them. > >The above statement is not quite true. If you design your program with all of >your eval functions stored in a file on the hard-disk and load them each time >you run your program, then the program can modify its own eval without having to >recompile its self. This is really quite easily to implement, but I do not know >of any programs that are doing it as of now. After a few hundred (thousand?) >games then the numbers can be hard coded if you want. That is only a part of "learning". If, by "learning" you mean only adjusting existing evaluation terms, then what you describe will do the trick, for whatever you can get from it. However, what about _new terms_? IE you don't know anything about a pawn majority, but you keep getting beat because your opponent has one that turns into a passer late in the game. How do you learn your way around _that_? Existing term modification won't fill the bill. You need at least one (or more) new term. Adding something new is suddenly not so easy... >Bill > > >To my knowledge, the most radical approach in this direction has been done >>in Checkers (Fogel et.al.). They used a neural network as evaluation function, a >>fixed 6-ply search with quiescence extension, and didn't write anything in the >>program except for the rules of the game, and a basic material evaluation (just >>counting pieces). Then they let the program play itself for about a year, and >>then let it lose in the GameZone. There it achieved a 2000 rating (comparable to >>chess ratings). However, it cannot compete with top checkers programs with >>hand-tuned evaluation functions. >> >>Now that I think of it, there was a chess program a while back that started on a >>server with a ~1700 rating, and brought that to a ~2200 rating using temporal >>differences. (was it KnightCap?) >> >>So as far as your question about current chess programs go, I don't know. What's >>nice for the average user is the following: When I got my first chess computer >>(some sort of small portable Mephisto), I realized that it screwed up some line >>of the Berlin Ruy-Lopez. Hence I found a way to gain substantial material in >>about 8 moves. It was fun to beat the computer, but only the first 4 or 5 times. >>If opening book learning is done properly, then after a while the computer would >>start playing something else, and give me something new to think about. That >>would be fun. (Incidently, a cousin of mine, still has some handheld computer, >>and still hopes that at some point it will learn to avoid such lines...) So I >>think this is really a plus. Moreover, it makes it more difficult for computer >>opponents to prepare something tailored against a given book. >> >>So, in conclusion, I think that learning is possible, and that it makes the >>programs more interesting. >> >>Sven.
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.