Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Chess, Backgammon and Neural Nets (NN)

Author: Jay Scott

Date: 07:38:50 08/21/98

Go up one level in this thread


On August 20, 1998 at 12:23:33, Torsten Schoop wrote:
>does anyone know why NN backgammon programs play very well (TD-Gammon,
>JellyFish, Snowie)but the NN chess playing programs do not?

No, nobody knows. Or if they do they haven't convinced everybody else. :-)

My favorite theory is inspired by the analysis of what kind of
backgammon positions the neural net programs are less good at playing.
They're what backgammon players call "technical positions", where
the choice of best move depends on calculation of the probabilities
of future events. In a typical backgammon position, the dice rolls
"smear out" what chess players would think of as the tactical
possibilities of the position, and the neural network's estimate
of the relative probabilities of events is excellent. In a technical
position, which is comparatively rare, expert humans can calculate the
probabilities closely, but the neural network still uses its everyday
estimate, which is not as accurate.

Now compare: in chess, most positions are "technical" positions
that depend on exact calculation! So you shouldn't expect a network's
estimate to be particularly accurate.

I have a brief discussion of it on my Machine Learning in Games site,
on this page:

  http://forum.swarthmore.edu/~jay/learn-game/systems/gammon.html

Some other theories are, briefly:

- Chess is more complicated.
- Not as much work has gone into good neural network chess evaluation.
- We don't happen to have discovered "good" neural network input
features for chess; they might be different from good traditional
evaluation features.
- We haven't figured out the right way to train a net to produce
the most useful evaluation function.

  Jay



This page took 0.01 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.