[Bug-gnubg] Training of the crashed neural net. What would help?

[Øystein Johansen]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

There's a posting at GammOnLine which shows a big blunder of GNU
Backgammon in a position which comes under the CRASHED category.

The position is:
Position ID: DwAEoG3tABAJAA Match ID: UQkAANAAoAAA

GNU Backgammon needs more training in the crashed net. The question is,
how can this be done.

The first thing I'm thinking of is to start more TD trainig of a set of
crashed positions. Will that help? Or is these positions so hard for the
neural net to play that the weights won't converge?

Can we change some of the inputs for this position class? Maybe it can
improve if we use more inputs on the checkers bourne off. Maybe one
input node for each checker off (like the race net) would improve it?
(The current crashed net uses 3 nodes to represent the checkers off.) Is
there a input for loose checkers?

Can the strength of the playing be improved by splitting the crashed net
into CRASHED and CONTAINMENT? Will that improve it?

Can someone please comment? Joseph?

- -Øystein


-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (MingW32)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFBYm1V6kDTFPhwyqYRAszoAJ9nbtDXMORCGy9z/6GKAs9IBBEGcACfe/eY
RzkcwI8My3k/2QIby1qIi+8=
=lo1A
-----END PGP SIGNATURE-----