Re: computer-go: abstract info and neural nets

[Erik van der Werf <E.vanderWerf@xxxxxxxxxxxxx>]

The network can in principle learn things like the counting of
liberties. However this is quite a difficult learning task and you may
want to focus the effort more on learning to detect features which are
not so easily handcoded. On the other hand, maybe some of the features
humans choose to encode are not needed because the net would suffice
with cheaper internal approximations. 

Bottom line: Always try to keep the dimensionality of the featurespace
low, but if a feature really adds somthing to the performance you will
have to include it somehow...

Erik


Julian Churchill wrote:
> 
>  Hi,
> 
>  I have been puzzling over something for a few days now and wonder if anyone
> has any ideas or opinions on it.
>  I am using neural networks as part of a Go playing program I am writing and
> so far have only fed the nets with basic information, i.e. the board
> contents. I have noticed that other neural net programs (NeuroGo and Nicol
> Schraudolph et al.) used much more complex approaches, inputting information
> calculated about strings and groups, such as number of liberties and
> connection possiblities.
>  I was hoping, perhaps naively, that the neural network might somehow
> 'discover' the relevance of such things and the importance of them be
> reflected in the weight adjustments and so it would be unnecessary to
> calculate all that extra, possibly complicated, data. So do you think it's
> better to let the network absorb these concepts, liberties, eyes etc.., (and
> is it even possible) or would it be more useful to a Go playing program to
> have some expert knowledge calculated previously and just fed straight into
> the net?
> 
>  I would be delighted to hear any responses.
> 
>  Cheers,
>  Jules
> 
> _________________________________________________________
> Do You Yahoo!?
> Get your free @yahoo.com address at http://mail.yahoo.com