Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Symbolic: Status report 2005.08.13

Author: Steven Edwards

Date: 16:25:13 08/16/05

Go up one level in this thread


On August 13, 2005 at 16:41:00, Michael Yee wrote:
>On August 13, 2005 at 13:52:53, Steven Edwards wrote:
>[snip]

>>How might it be possible to partially automate the process of writing
>>productions?
>
>Have you considered genetic programming? Overall, producing good productions
>automatically sounds like a very hard discrete optimization problem. But maybe
>since you're trying to mimic human play anyway, you can avoid self-play and just
>train on human databases (which could save time).

GA was good for the mate attack subproject, but I have doubts about its ability
in the production context.  Analyzing human games will be very important.

>>Can the productions be automatically evaluated for merit?  If so, is it
>>reasonable to have a feedback loop to implement a simple form of program self
>>adaptation?
>
>What about forming a database of when the productions led to wins/losses/draws
>and also keeping track of a vector of (real-valued) position features present
>when these results occurred (e.g., % phase of game, piece activity, pawn
>structure, etc.)? Then you could fit a probability of success as a function of
>this feature vector.

This is certainly a possibility.



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.