RE: [computer-go] Pattern matching - example play

[Vincent Diepeveen <diep@xxxxxxxxx>]

At 18:30 4-12-2004 -0200, Mark Boon wrote:
>
>
>> -----Original Message-----
>> From: computer-go-bounces@xxxxxxxxxxxxxxxxx
>> [mailto:computer-go-bounces@xxxxxxxxxxxxxxxxx]On Behalf Of Vincent
>> Diepeveen
>> Sent: Saturday, December 04, 2004 17:43
>> To: computer-go; computer-go
>> Subject: RE: [computer-go] Pattern matching - example play
>>
>> If you make a small neural net you could just as well write the pattern
>> knowledge yourself.
>>
>> If you create a big neural net of many layers, then training it
>> is going to
>> take too long.
>>
>> That's the problem of neural nets.
>
>That's a possibility, as I said I don't have a lot of experience with Neural
>Nets. I don't know to what point they can do some 'magic' by achieving a
>weighting system that would be hard to achieve by hard-coding. But I know
>that the coding is definitely hard :) and error-prone.

Neural networks are fun to toy with as they serve 1 important purpose:

When a pattern is easy to recognize, as the pattern itself is very rudely
defined, you make a neural network and let the ann with automatic learning
figure out the pattern to recognize.

That simply removes all that hard work for you to roughly handcode the
pattern in a crude way.

Sometimes handcoding the pattern is pretty difficult anyway as it is
depending upon the user using it and the circumstance under which it gets
applied.

In such a case an ANN is very useful.

Good example is predicting presidential elections:
  a) it is allowed to estimate it roughly
  b) whatever choice you pick, people believe it
  c) it avoids you from retuning all those parameters each time
  d) because an ANN predicts it your company doesn't look like biassed

D is very important. here in netherlands i know 1 director of a company who
also is doing opinion polls and his results are very biassed towards a
certain political party P. When asked straight to the man the question why
his polls are more to the leftside always than to the rightside measured
and have a very consequent tendency to be more leftwing than rightwing,
especially favouring party P. His answer was: "well when we poll it is what
they say what we write down here, but when they are in front of that voting
machine somehow they push a button more to the rightwing".

Very political explanation from him.

However this director is a very important and influenceal member of
political party P.

So in USA solving this problem using a neural network was pretty clever.

"The machine predicted it"

Looks rather more objective than what happens here.
Note neural network expert creating that ANN = Greg Young.

Yet you read another problem. An ANN is a very crude and primitive thing.
To judge go positions you need something very fine grained and sophisticated.

You can solve the fine grainedness by making the neural network at least
quadratic bigger and let it learn for another zillion years longer at 
hardware we have in a zillion years from now. So theoretical the proble is
solvable. Practical. Nah, too much computing time.

However that's exactly what ANN's were intended for. To solve things by
throwing computing power at it while you sit in your lazy chair watching
the news.

Yet that learning time is the real problem for finegrained gameplaying
programs.

Vincent

>If NNs work for statistical games like backgammon, then I wouldn't want to
>rule out its use for Go. I know someone who calls Go (tongue-in-cheek) a
>game of chance. Although it sounds laughable at first, there's more than a
>grain of truth in it. Go is so complex that even for the strongest players
>there are a lot of uncertainties how the game is going to develop. It's a
>game of perfect information in theory only. 'Good shape' for example, is
>often (not always) nothing more than saying: empirically, in the past, it
>has shown making these shapes cause problems later in the game less
>frequently than 'bad shape'. Often human players agonize over making an
>empty triangle. Their evaluation says there's immediate gain in playing it,
>but their experience has given them statistical data indicating it could
>still be a bad choice. When does one precede over the other is fuzzy to say
>the least. And I thought NNs would be good at coping with this kind of
>fuzziness. But it's just a guess.

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