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

[Vincent Diepeveen <diep@xxxxxxxxx>]

At 09:57 4-12-2004 -0200, Mark Boon wrote:
>
>
>> -----Original Message-----
>> From: computer-go-bounces@xxxxxxxxxxxxxxxxx
>> [mailto:computer-go-bounces@xxxxxxxxxxxxxxxxx]On Behalf Of Heikki Levanto
>> Sent: Saturday, December 04, 2004 9:18
>> To: computer-go
>> Subject: Re: [computer-go] Pattern matching - example play
>>
>>
>> On Fri, Dec 03, 2004 at 10:27:58PM +0100, Vincent Diepeveen wrote:
>> >
>> > >I would not completely write off higher-level planning and neural nets,
>> > >and other fancy theories. Many of them have shown their values in
>> >
>> > We can have lengthy discussions, but the majority of ANN top researchers
>> > agree with me here that for game playing ANN is completely useless.
>>
>> Well, as far as I know, neural nets work quite well for backgammon. I
>> admit it s a very different game, but shows that ANNS are not
>> *completely* useless for *game* *playing*.
>>
>> I agree with you, that if you just feed the board position into an ANN,
>> you can at best train it to recognize similar games, and you will not
>> anywhere. But there are other ways to use ANNs.
>>
>> For example, if you feed it the number of groups, the number of their
>> liberites, and so on, it should be relatively easy for it to learn that
>> a position where groups are connected is a better one than a situation
>> with many small isolated groups. And if you feed it the results of some
>> influence calculations, it should easily learn that the one that has
>> most territory is often ahead. Both of these are easier to program in
>> hand, I admit. But the same ANN could also learn to balance these
>> separate considerations, and a few more, and come up witha  decent
>> evaluation function.
>
>Although I'm a beginner when it comes to neural-nets I can see lot of
>interesting potential for neural nets for problems that so far have eluded
>current Go programs. Recognising and evaluating aji is one. An important

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. 

Human brain gets underestimated too much. Note that human brain works quite
different from ANN. 

Yet example. De Groot has estimated back in the 50s that professional
player know at least a 100000 patterns. In Go that will be probably even more.

Most such patterns are pretty complex.

Storing that in a neural net will require litterary 500000 neurons.

Try to train a complex 500k neural network!

It has loops too!



>concept that I don't think any program handles well, if at all (Goliath has
>a half-hearted attempt in it.) Properly estimating the effect on the score
>of having one or more weak groups is another. This goes in general for
>weaknesses in a position and is closely related to aji. If you can't
>recognise aji, then a program is always immediately going to play aji-keshi
>if it results in a small short-term gain in the evaluation. For brute-force
>approaches this will be another serious road-block.
>
>"The worst move is a forcing-move that is always sente." a wise lesson by
>Otake Hideo. Really understanding this makes almost any but the strongest
>amateurs a stone stronger.
>
>_______________________________________________
>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/