RE: [Bug-gnubg] Two patch sets - one for Joseph Heled

[David Montgomery]

> I think there's a lot more performance to be gained, but it may be
> harder work - the biggest time consumers are the setting up of the
> neural net inputs and the evaluation of the net itself, neither of
> which are trivial to change.

Here's one idea for speeding up network propagation.  

Since most evaluations are of a position very similar to the 
previous evaluation (e.g., for checker plays), the inputs are 
mostly similar as well.  So it may be faster to just update the 
net with the *differences* in the inputs, rather than starting 
fresh each time.  One could do this either by looking at the 
diffs between this position and the last evaluated, or the do all 
checker plays by looking at the diff from the starting position.

If anyone wants to look into using this idea, let me know and I'll 
send you code that implements it for the raw board encoding.

David