Re: [computer-go] question regarding Hydra Chess PC computer

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

----- Original Message ----- 
From: "Vincent Diepeveen" <diep@xxxxxxxxxxxxxxxxx>

> I'm sure in go and chess and any board playing game you'll see similar
> things. professional players are the most horrible source to 'improve' 
> ones
> evaluation function. They do not understand how to make an evaluation
> function and will never understand what a computer is and how it works.
I agree. Last year, a 1p player offered unlimited assistence.
We talked a lot at first, brainstorming, me asking questions etc.
I came to the conclusion that at least the coming years I would have no use 
for a pro player (imagine, I do not play Go myself!).

On the verge of being deemed immodest, I even think that pro advice is not 
so useful *at all* in Go. Simply because Go appears much "fuzzier" than 
chess. In Go, it seems to pay off greatly to simply do large-scale 
statistical analysis of patterns and "prove" that chains can be connected 
(patterns, search) instead of relying upon heuristical equivalents of 
proverbs and pro's gut feelings.

There seems much more progress possible for a computer in Go compared to 
humans, as for a computer in chess, compared to humans, and I suspect that 
pro advice in comp. Go is even much less useful as in comp. chess.

Also I think that statistical analysis in Go is more promising than in Chess 
due to the highly statical nature of Go positions. When you are able to 
measure metrics like connectivity, territory, moyo, influence and have a 
decent tactical module, you come a long way.

I have experimented with a connectivity estimator. How it works: Examine 
500,000 games and record every occurence of two or more chains in eachother 
neighborhood.

Record tens of million of normalized patterns and keep a score for each 
pattern: "becomes connected" and "does not become connected". Next time you 
encounter a pattern, you can say something about the statistical likelihood 
that the chains will become connected or are connectable. And if the pattern 
has never been seen, it is likely non-connectable.

This I managed to make work quite well, but the results were a bit 
disappointing. Because chains can connect somewhere else than at the 
pattern's location. Chains often span quite an area. And connections 
sometimes depend on ko-fights elsewhere and other things I don't know about 
:)  Another problem was that the database was 64 MB and added to the 80 MB 
of the pattern database, it was not commercially viable. 
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/