Re: [computer-go] Pattern matching - example play

[Imran Ghory <imran@xxxxxxxxxxxxxxx>]

On Sat, 4 Dec 2004, Vincent Diepeveen wrote:

> At 22:20 3-12-2004 -0500, Jeffrey Rainy wrote:
> >Hi,
> >
> >> > We can have lengthy discussions, but the majority of ANN top researchers
> >> > agree with me here that for game playing ANN is completely useless.
> >
> >Ok, I'll bite on this one.
> >
> >Any computable function can be computed by a given (possibly large) ANN. Any
>
> Yes i was waiting for someone to bite.
>
> Practical limitations avoid that. If you use an ANN that *theoretical* can
> do the same like patterns that are in a pro program, then the time to train
> your ANN in a correct way is like 10^200 possibilities. Somewhere in that
> region.

However improvements in things like board encoding, training algorithms
and raw computational power are bringing many more games into
practicallity.

Obviously you haven't seen the Temporal Coherence Algorithm, only limited
work has been done with it so far, but the results have been very
positive. It has been used to train a neural network which takes as inputs
(human programmed) feature detectors, when you just use a single layer
neural network you're essentially just producing weightings for the
different features. Research in games such as Chess and Shogi has found
this algorithm to be the best way discovered yet for combining feature
detectors to produce a score for a board position.


BTW. If you would like to know more about the application of TD algorithms
and neural networks to games see my paper at,

http://www.cs.bris.ac.uk/Publications/pub_info.jsp?id=2000100

Particularly section 9 might interest you.

Imran
-- 
http://bits.bris.ac.uk/imran
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/