Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Temporal Difference

Author: Jay Scott

Date: 10:39:22 01/07/01

Go up one level in this thread


On January 07, 2001 at 05:35:32, Gian-Carlo Pascutto wrote:
>One of the papers I have read (I think the KnightCap one) had
>a comparisaion with SAL, and the authors concluded that only
>learning from losses caused the eval function to grade every
>position as a loss.
>
>You're not seeing that effect?

That can happen because it's a correct solution: An evaluator
which believes that it always loses will, against most opponents,
always lose. It learns to evaluate every position as a loss and
ends up playing randomly.

Training on losses only might cause this problem, but I don't think
it necessarily does. The degenerate solution is not the only one.

  Jay



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.