RE: [Bug-gnubg] New positions for training database

[Ian Shaw]

Ian Shaw wrote on 24 July 2006 18:32

> 
> Ian Shaw [mailto:address@hidden 
> 
> > Has anyone tried going back to TD training after a period of 
> > supervised training? Perhaps now that gnubg has got over 
> its sticking 
> > point it can learn more from TD. Perhaps there are 
> subtleties it can 
> > now discover now that it is already an expert.
> 
> Does the command Train TD start from the current set of 
> weights, or does it initialise all the weights to random?
> 
> If the former, I might try setting it going while I'm on 
> holiday to see what I end up with when I get back.
> 
> Is there a command to save the new weights, or is it written 
> to disk automatically?


I've run a short test. I left gnubg doing td training for a couple of
days. I stopped it and used "save weights" to generate gnubg.weights.
I deleted the original gnubg.wd file and played a test game. Gnubg
appears to be playing a decent game, so I surmise that it is starting
with the original weights.

"Show engine" gives the following output.

 * Contact neural network evaluator:
   - version 0.15, 250 inputs, 128 hidden unitstrained on 68375431
positions.

 * Crashed neural network evaluator:
   - version 0.15, 250 inputs, 128 hidden unitstrained on 68375431
positions.

 * Race neural network evaluator:
   - version 0.15, 214 inputs, 128 hidden unitstrained on 5280099
positions.

You don't normally get the "trained on 68375431 positions." part, which
seems to indicate something has changed.

It is suspicious that the contact and crashed networks are both
identifying 68375431 positions. There should be many more contact
positions reached than crashed.

Is gnubg skipping the inappropriate position but still incrementing the
count, or is it training both nets on each position?

-- Ian