Author: James Swafford
Date: 12:38:49 07/31/02
Go up one level in this thread
On July 31, 2002 at 01:00:47, Russell Reagan wrote: >On July 30, 2002 at 22:43:36, James Swafford wrote: > >>A natural follow up question (which I also asked) is -- then why isn't >>everyone doing it?? > >My first thought on this is that the top chess engines already have their >evaluation weights tuned very well. Perhaps not perfect. I think they are >putting their efforts into more beneficial things. For example, if your >evaluation function weights are 99% correct, then it's more beneficial to work >on a new pruning technique or add new evaluation factors, but tuning that almost >perfect evaluation function isn't going to produce any significant increases in >playing strength. > >>Knightcap was strong, but it's >>definitely not in the top tier. > >My thought on this is that just because an evaluation function's weights are >tuned to 100% perfection, it doesn't mean the engine will be strong. Maybe you >are evaluating the wrong things to begin with. If you only have material, >mobility, and king safety, then I suspect there is only so much tuning you can >do, and eventually you get optimal weights for each of those evaluation >parameters. Just because those are perfect, it doesn't mean the engine will be >among the top. If that program has a branching factor of 4, it's not likely that >it will ever compete with Fritz, Tiger, Shredder, Junior, etc., all of which >have much lower branching factors. > >I think that is the main reason why you don't see (or hear) about TD in computer >chess very much. There are other things that will benefit the engine more. I >would think that even when a programmer reaches the end of his "to do" list, >that he would probably find it more beneficial to try and create completely new >pruning methods from scratch than to spend time working on implementing TD >evaluation tuning. Seems from reading this you will fall into the "search" camp. Yeah, that's great and all, but as hardware improves, and if you believe that at some point there are diminishing returns from search (and I do), then you will need more than search heuristics. You will need more knowledge. You're right about needing the right knowledge... then worry about tuning it. The nice thing about TD methods is that it relieves some of that burden. Any patzer (like me) knows that it's important to have rooks on open files. How much it's worth is an entirely different question, especially when you have tons of terms. I think the next _big_ step is computer chess will come from the automatic discovery of new knowledge... both the terms and their weights. -- James > >Russell
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.