Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Evaluation at start versus eval at node. Why not mix them?

Author: Dan Homan

Date: 15:58:09 06/11/99

Go up one level in this thread


On June 11, 1999 at 17:43:59, Jon Dart wrote:

>On June 11, 1999 at 10:39:13, Dan Homan wrote:
>
>
>>I've been thinking about this alot recently.  One thing that occurred
>>to me is that this might be an effecient way to create a neural net based
>>program.
>>
>>Say the neural net (nn) is responsible for filling a piece-square table...
>>Then the nn could operate either once, at the root of the search, or
>>only selected times during the search.  This would largely overcome
>>the major drawback of nn evaluations which is that they are slow.
>>
>>Then we could have a TD based learning program that also learns
>>evaluation features!
>>
>>I've thought alot about building my next chess program on this
>>idea....  The one major drawback that I see is that there will be
>>a relatively large number of parameters for learning to adjust.
>>The other drawback is that I really know nothing useful about
>>programming neural nets.  :)
>>
>> - Dan
>
>KnightCap (http://samba.anu.edu.au/KnightCap/) has learning of
>its evaluation parameters (as well as book learning) .. it
>appears to work very well. But I don't believe it does non-leaf
>evals.

I've read the knightcap papers.  I was thinking it might be interesting
to go a step beyond learning values for pre-defined evaluations features
and actually learn new features not previously defined.

Neural Nets will allow this, but they are slow.... so I was thinking
of using a neural net to fill piece-square tables at the start of the
search or at well-defined points in the search.  Maybe this wouldn't
work, but it is interesting to think about.

 - Dan
>
>--Jon



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.