Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Neural Networks Recognise Patterns!

Author: ujecrh

Date: 16:26:37 06/29/00

Go up one level in this thread



>
>I don't understand what you mean by "pattern."
>
>If you mean something that humans can't scientifically articulate, like "king
>attack," then a NN might be a good idea. (...although a NN that can identify
>king attacks might as well just play chess by itself...)
>
>But there are a tremendous number of patterns which can be recognized with 100%
>accuracy with simple procedural logic (e.g., "doubled pawns"). Using a NN to
>detect doubled pawns would be extremely stupid.
>
>-Tom

Yes, maybe we all put a different meaning behind "chess patterns". For me, they
represent characteristics of a position that are suceptible to allow a special
treatment.

For example in a given (simplified) pattern (white: Bd3, Nf3,Qd1; Black: o-o
with nf6 and no bishop or queen on d8-h4 diagonal) an average player might
immediately recognise (without calculation) the manoeuvre Bxh7,Qh5,Ng5 and then
put Bxh7 in top of his search. Of course it is wrong in most cases but in some
patterns it proves to be efficient.

This example is obviously not really good because it is often easy to find with
brute force search (I just tried to keep your idea of king attack) but such
schemas that allows to think at more long terms objectives are what I personally
mean by patterns. Particularly in endgames where a particular pattern allow to
apply a particular rule without much tactics involved.

A good example was a game between Anand and Genius 3 in London some years ago
(the round after Genius beat kasparov at PCA championship): if I remember
correctly, genius king took a pawn on the h file which was an obvious mistake.
Masters commented this move without any tactics just by knowing that was wrong,
they had the pattern, genius did not.

Then we can distinguish two kinds of patterns. Those that can be recognised by
pure logic (algorithm) and those that are only a matter of experience and
subjectivity. The latter might be a good NN application.

Maybe a program based only on NN would be wrong but splitting its knowledge
between standard methods and NN could be interesting.

Well, this is the way I see it but it may be a complete nonsense for others.

Ujecrh



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.