RE: computer-go: Perl Module for next move.

["Mark Boon" <tesuji@xxxxxxxxx>]

> -----Original Message-----
> From: owner-computer-go@xxxxxxxxxxxxxxxxx
> [mailto:owner-computer-go@xxxxxxxxxxxxxxxxx]On Behalf Of heikki@xxxxxxxxxxxxxxxxx
> Sent: Tuesday, June 05, 2001 8:28 PM
> To: computer-go@xxxxxxxxxxxxxxxxx
> Subject: Re: computer-go: Perl Module for next move.
>
>
> I remain quite sceptical about the possibility of such a network ever
> learning to deal with even simple tactical considerations, like reading a
> ladder. But I wish you best of luck, and hope you'll keep us posted.
>

I'm very skeptical as well. But not so much because of tactical
considerations. In general I think in order for a neural network to be able
to learn Go this way, it will have to learn new technigues and concepts by
itself. Tactical technigues are just part of that. But I have doubts you'll
be able to make a neural network big enough, and have it learn fast enough,
without a massive paralel-computing device.

I'm not into neural nets at all, so I don't know its current state, but are
there any known projects where a neural net has actually resulted into
developing the concept of reading ahead? Instead of taking on Go, wouldn't
it be better to develop a neural network for a simpler game like
connect-four to see if it can learn to play the game well? I think Go has a
too big problem and concept space to learn from scratch by a neural net just
like that. Not with todays hardware. But if you can make a neural net learn
to play connect-four without putting in any explicit knowledge at the start
I would be very impressed.

In a previous post a checkers project was said to use neural nets
successfully, but I'm wondering how much was done by a fast mini-max program
which used the neural net in some way (as a move-selector for example)
instead of having the net learn the mini-max strategy as well. I somehow
feel the latter is necessary for neural networks to be successful in Go.

    Mark