Re: current development

[Philippe Michel]

On Thu, Dec 05, 2019 at 11:31:33AM -0500, Timothy Y. Chow wrote:

> As I recall, in the past, people have tried specifically training neural 
> nets on positions like these, as well as "snake" positions where you have 
> to roll a prime for a long distance, and the problem was that it seemed to 
> degrade performance on other types of positions.

A few hundreds of such positions were added in the training set for 
gnubg 1.x. Starting not from pure snake positions, but a relatively 
likely forward formation + 6 to 10 checkers far in the outfield playing 
against one or two stragglers, then adding the 2 or 3 best moves 
according to the previous net and the 2 or 3 that looked most plausible 
to me (they often did not overlap...) to the training set.

It made a substantial difference with no obvious degradation on more 
standard positions. An example doesn't prove anything, but in your 
position, getting :
    1. Cubeful 2-ply    18/16 13/7*                  Eq.: +0.174
    2. Cubeful 2-ply    10/4 6/4                     Eq.: +0.155 (-0.019)
    3. Cubeful 2-ply    18/16 10/4                   Eq.: +0.136 (-0.037)
is hardly ridiculous. Either of the first two could easily be better 
that the double hit.