Author: Don Dailey
Date: 10:09:13 05/15/98
Go up one level in this thread
On May 15, 1998 at 05:01:05, Graham Laight wrote: > >If we're going to talk "meta-programs", why not go the whole way and >talk about programs that WRITE chess programs? > >Wouldn't that be more useful? > >If someone would write a program that would create chess programs (CPs) >that could apply different knowledge in different types of position, >then automatically test these CPs against each other, wouldn't we have >something that we've forgotten we all wanted - an automatic chess >learning machine? > >Wouldn't this pass the test of artificial intelligence, make everyone >happy, and generally be a good thing to do? Great idea. Go ahead and show us how. Ok, I'm being sarcastic here sorry! This is clearly easier said than done. A lot of my own chess code was written by programs I wrote to write them! But nothing really interesting, I do this to avoid errors from tedious code. I think all the programmers write little programs that write their table declarations, but of course this is not really what you're talking about. I have done experiments with genetic algorithms that adjust evaluation function parameters by playing thousands of games. This is not quite the same thing either but it's similar. I know John Stanback has played around with this too. I'll bet many of us have. For anyone interested in this, I think it's actually fairly promising if correctly engineered. One very big problem which I believe is a BIG problem in general for all of us, is the "fitness function." If you cannot very quickly measure which versions are better, you have a serious problem with GA's. GA's often depend on slow incremental progress and that's a big problem. If this problem could be solved, then every programmer would benefit enormously, GA's or not. I see it as perhaps the biggest obstacle to overcome. How can you tell if a minor improvement helped the program? It's impossible (from a practical standpoint) to measure 5 or 10 rating points. But getting this a few times can put your program in a new class. - Don
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.