RE: computer-go: abstract info and neural nets

["Staelin, Carl" <staelin@xxxxxxxxxxxxxxx>]

> I am in favor of full board NN evaluation.
> 
> I found bunch of demo GO programs like EzGO, 
> HandTalk, ManyFaces, Goliath, and TurboGO, 
> etc. on 9x9 board, and have them played to 
> each other awhile to collect about hundread 
> games. Then, I use NN to get an initial full 
> board evaluation function. My program 
> (ForeverBlack) then uses the NN to play with
> itself and other programs, and the new game 
> records are used for further learning.
> 
> After 400+ games, I found ForeverBlack started 
> making living group, connections, and kills. 
> The programs behaves no worse than influence-only
> move (such as the influence function used by 
> HandTalk), and first 10~20 moves of a game 
> usually make sense.
> 
> With the game records grow, I constantly needs 
> to increase hidden nodes of ForeverBlack's NN, 
> that's painfull. Because the training time 
> becomes longer and longer. My 800mhz computer 
> can not get weights converged in a full half
> week's learn.
> 
> Can someone point to me literature or articles 
> about incremental construction of NN for 
> regression (piecewise continous function regression)?

One paper I would suggest you read is Xin Yao's
paper "Evolving artificial neural networks" in
Proceedings of the IEEE, 87(9):1423-1447, Sept.
1999, which is available online at:
http://www.cs.adfa.edu.au/pub/xin/yao_ie3proc_online.ps.gz
He uses an evolutionary system to both prune and
grow networks incrementally, and you might use a
similar approach to simply grow the networks.  He
uses a generalized feedforward network (GFF)
architecture, which may prove useful.  I think the
biggest question is where one should (usefully)
add new nodes.

Another question is how are you training, and 
what is the network supposed to be learning?
Are you using a reinforcement learning-type
approach where the network learns to estimate
the "value" of a board position, or are you
doing something different?  I would be very
interested in hearing what inputs you are giving
the network, and what outputs you are asking it
to learn.  I would also be interested in how
you are training the network, e.g., via 
reinforcement learning or supervised learning.

If you are using a somewhat straightforward
batch-type approach with traditional back-
propagation momentum-based learning, then I 
think you might usefully switch optimization
algorithms to use a quasi-Newton training
algorithm such as BFGS, which is *far* more
efficient.  If your networks get too large,
then you might switch to a limited-memory
BFGS optimization algorithm.

Again, depending on how you are training the
networks, you might look at a completely 
different network training approach, such as
NEAT which uses an evolutionary approach to
choose both the architecture and set the 
connection weights.  You might look at Kenneth
Stanley's paper "Evolving neural networks through
augmenting topologies", which is available at:
http://www.cs.utexas.edu/users/nn/pages/publications/abstracts.html#stanley.
utcstr01.ps.gz


Cheers,

Carl

_________________________________________________
[(hp)]	Carl Staelin
	Senior Research Scientist
	Hewlett-Packard Laboratories
	Technion City
	Haifa, 32000
	ISRAEL
	+972(4)823-1237x221	+972(4)822-0407 fax
	staelin@xxxxxxxxxxxxxxxxx
_______http://www.hpl.hp.com/personal/Carl_Staelin_______