Author: James Swafford
Date: 16:49:59 08/05/02
Go up one level in this thread
On August 05, 2002 at 18:24:51, Sune Fischer wrote: >On August 05, 2002 at 17:43:51, James Swafford wrote: > >>>>>Perhaps once you are convinced you are evaluating most of the really important >>>>>things, this might work. I'm a long way from that point myself... >>> >>>It sounds to me like a strange problem, I think I will have to see that for >>>myself to believe it. I would expect the temporal difference to be 0 if there is >>>a knowledge term missing, so i don't know what could be going on. >> >>No... if you are missing something important, other parts of the eval will >>attempt to "compensate"; terms will be adjusted too high or too low. It's >>really important to get a pretty good set of terms for TD to work. >> >>An interesting observation if you think about it... suppose you know pretty >>well what the weights should be. Then remove an important term, and see >>what happens to other terms. The terms themselves may suggests what is >>missing. (Schaeffer told me that, I've been thinking about it since.) >> > >Ah ok, I will take your word for it then :) >But if there is something missing, and other terms try and compensate, is that >really a bad thing? >It seems to me that you _need_ to compensate somehow for this lack of knowledge. >The algorithm will do what it can, could be the weights will look strange to us, >but if the result is better play then what does it matter? > Right, if you're missing something, then the best you can do is to optimize for _those_ weights. Whatever values you end up with, they should produce the best results for that vector of terms. I think some people have it ingrained in their head that certain weights should have certain values, ignoring what it is the weights are doing for you ( contributing to the selection of good moves ). -- James >-S.
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.