Re: [computer-go] Weights are very important

["Nicol N. Schraudolph" <compgo@xxxxxxxxxxxxxxx>]

Vincent,

> >If you know a priori that the value should be between 1 and 40, why
> >on earth would you not initialize your neural net within that range
> >and let learning refine it from there?
> 
> For a very simple reason. If you believe that NN's work, it should figure
> that out itself of course. If you don't believe in it, you don't use them
> at all.

you seem to have fallen in and out with a rather fundamentalist sect
of NN believers!  I for one don't believe in NNs by your definition,
even though I make my living with them and consider them useful tools.

> If your only knowledge is that you lose games, then i do not see how to
> independantly tune 1 parameter.

Fortunately in go we have far richer information about the game's outcome,
namely who owns which points at the end.  I agree that you're not going to
get very far in go with an ML approach that just uses win/loss.

> Several researchers amazingly claim that it is possible to tune sets of
> parameters (usually it appears they do that by randomly changing a number
> of parameters) at the same time.

Sure that's possible, using either gradients or simultaneous perturbation
approaches (e.g., http://www.jhuapl.edu/SPSA/).  The millions of NN weights
I've mentioned were all adjusted simultaneously so as to minimize a single
number (the loss function).  It's humans who get confused by simultaneous
changes to multiple parameters, the computers can keep track just fine.

Now, I agree with you that NNs may well be next to useless for the
specific task you want them for.  The difficulties you describe - poor
error signal, expensive data collection, inavailability of gradients,
etc. - are just what I mean by "poor match between frameworks".  NNs
are no magic bullets, they only work well under certain conditions.
Markus Enzensberger has spent a lot of thought and effort on how to put
go knowledge into a NN-friendly framework, and his NeuroGo seems to be
doing quite well.  What you get out is only as good as what you put in,
whether you use machine learning or not.

> Happy tuning.

Thanks, and the same to you.  Best,

- nic

-- 
    Dr. Nicol N. Schraudolph                 http://n.schraudolph.org/
    Steinwiesstr. 32                         mobile:  +41-76-585-3877
    CH-8032 Zurich, Switzerland                 tel:      -1-251-3661

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://computer-go.org/mailman/listinfo/computer-go