Re: Fuseki and Joseki Database with Neural Network

[David Mechner <mechner@xxxxxxxxxxxxxxxxx>]

Magnus wrote: 

(To explain why datamodels is better)
"I think that ANN has the potential to discriminate those moves that 
seems to be alternatives to the human eye from those that are never 
considered at all. But an ordinary ANN however properly trained, will 
not be able to make the correct choice between these alternatives."

I think this is an accurate evaluation of go ANNs to date, but I don't
think they've realized their full potential. I think the poor results of
go ANNs is partly a result of the choice of inputs; the low level of
representation of the state of the game. I think with suitably rich
inputs, ANNs could do much better than they have, and using a Joseki
database as (part of) a training set seems quite reasonable. 

-David
p.s. naturally this largely just shifts the burden of the problem to how
to generate "suitably rich" inputs.. you just can't escape :)