Re: [computer-go] Computer Go hardware

["Frank de Groot" <frank@xxxxxxxxxxxxx>]

From: "David Fotland" <fotland@xxxxxxxxxxxxxxxxx>

>I don't know anyone that's worked full time on computer go
>for very long.

Well I did but due to chronic illness I spent not much more actual time on
it than you did, per week. I sit behind the PC but nothing comes out of my
hands because instead of concentrating I get paranoid rages :) I would say I
spent about 0.5 actual year.
About half of that was GUI-work, like making toolbars dock better :)

I started a contract job on my birthday two days ago after having been at Go
for the past 2 years but I quit after one day. Embedded C/Assembly and linux
on a terminal can be OK but not when you're close to something marketable..
It's not that I can't use the money..


> Tuning the heuristics by hand is pretty tedious.

I try to avoid that.
It took me 2 years to figure out a way that will really help a lot in the
next stage.
I reached about 15% pro-prediction with my current pattern system.

I want to bring that to 20% or a bit more.
So I will use the pattern system as one input of a NN, but a lot more stuff
will go in.
Stuff that my pattern system has no grasp of. So when the NN is designed
well and when it gets the proper training, it should significantly
outperform the pattern system.

So my job is to say: "There is such a thing as patterns, pattern have a
min/max nr. of stones in them, patterns have sizes, patterns have values"
etc. and the pattern system learns them by itself and adjust their values
too.

Same with the NN of course but I will use that only to augment the pattern
system at the moment and will not yet improve the fact that it is near-blind
to tactical situations.

I'm about to start the NN experiments in a few days.
They look very promising, I did some preliminary work and some good moves
that were consistently rated very low by the pattern system (due to its
weaknesses) suddenly got into full focus by the NN.


>I don't think it is possible to make a strong program by extracting
millions
>of patterns from professional games and sorting through them automatically.

Me neither.
Neither will a strong program be produced by combining a pattern system with
a Neural Network, except when you manage to extract tactical info. It would
be nice but I agree that search is unavoidable.

I think that 4 ingredients cooked the right way will do the trick:

- pattern system
- neural network
- search
- genetic optimization of various parameters in and between the pattern
modules, search and NN.

All combinations may be neccessary, like using a part of the pattern system
in the first few ply of search etc.

But that is my plan.
I will take search very seriously, meaning I will spend years on it.
I see it as the core of a Go program that has a shot at becoming really
good.

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