Re: computer-go: Learning from existing games

[Markus Enzenberger <compgo@xxxxxxxxxxxxxxxxxxxxx>]

On Tuesday 14 January 2003 12:11, schraudo@xxxxxxxxxxxxxxxxx wrote:

> Agreed.  Markus Enzensberger has been doing very interesting work in
> that direction.  Markus, if you're listening - how is NeuroGo doing
> these days?

with about 1e5 weights, it can achieve > 45% wins against GnuGo.
The input has higher level features than the early versions, but still 
relatively simple ones compared to other Go programs.
Mostly the increase in playing strength come from a sophisticated 
network architecture. There is a diagram shown in these slides:
http://www.cs.ualberta.ca/~emarkus/nngo/nngo-2up.ps

The big problem is the training time. I found TD with self-played games 
superior to other training methods and it takes at least 100000 games 
for best results (several weeks or even months on a fast PC).
Algorithms for faster weight update are not necessarily helpful, because 
they decrease the exploration that is done with a given set of weights.

This is why I stopped my experiments with using GnuGo or Explorer for 
calculating input features, those programs are much too slow.

- Markus