Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Research idea? re: weight optimization

Author: Tom Kerrigan

Date: 04:44:55 12/29/01

Go up one level in this thread


On December 29, 2001 at 05:21:21, Rafael Andrist wrote:

>On December 29, 2001 at 05:11:26, Tom Kerrigan wrote:
>
>>Here's my idea.
>>
>>You have a position and you want your program to play a certain move (which it
>>presumably isn't playing). You run this algorithm:
>>
>>1. Search the position, get a PV. The evaluation of the last position of the PV
>>is eval(1).
>>2. Search only the move that you want your program to make, get a PV. This
>>end-point evaluation is eval(2).
>>3. Figure out which eval terms are different between eval(1) and eval(2).
>>Decrease the weights of all the different eval(1) terms slightly. Increase the
>>eval(2) terms slightly.
>>4. Repeat until the program plays the move you want.
>>
>>You could run this on lots of positions from GM games, to get your program to
>>play like a GM. (At least in some positions, heh.)
>>
>>Has this been done before? Are there any glaring problems with this idea? Does
>>anybody want to try this? If so, I'd like some credit for it. If not, I'll
>>probably get around to trying it sometime...
>>
>>-Tom
>
>You may be interested in the temporal difference learning algo by Richard Sutton
>which is implemented in the program KnighCap by Andrew Tridgell and Jonathan
>Baxter.
>http://www.syseng.anu.edu.au/lsg/knightcap.html

Thank you. They've done some terrific work. My idea is different from theirs,
though, I believe. To make broad generalizations, their goal is to make the eval
function a better predictor of the future, while my goal is to make it produce
known-good moves. Different means to the same end, presumably... :)

-Tom



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.