[Bug-gnubg] Re: Bug-gnubg Digest, Vol 89, Issue 11

[Frank Berger]

Hi Ian,

> Øystein Johansen and I have been messing around with training gnubg for a 
> while. We have managed to train a contact net that plays a little better than 
> the current net against both the benchmark and the current net itself.
> 
> The inputs and outputs are the same; the downside is that it far larger, 
> having 512 hidden nodes rather than the current 128. This may render it 
> unsuitable for use, since it will slow the evaluation speed considerably. 
> Given the current speed advantage already reportedly enjoyed by Extreme 
> Gammon, this may be too much of a penalty for what will surely be a modest 
> performance boost.

Have you measured the influence of the size of the hidden layer on the playing 
strength or was 512 a shot in the dark?

I've done that last year with BGB with 80,100,120,140,160,200(BGB current) and 
300 hidden neuron, where I trained all nets under similar circumstances. As one 
would expect, the playing strength increased, but there was no really 
significant result. The best value was with 160, but the differences above 140 
where insignificant and may be caused due to random influences.

Does anyone know the dimension/architecure of XG or Snowie?

ciao
Frank