Re: [computer-go] Pattern database

[Darren Cook <darren@xxxxxxxxx>]

> > This is my try to design a pattern database.
> > http://senseis.xmp.net/?PatternMatching
This is similar to the approach I used many years ago: i.e. a hash code that 
had a core and then added stones in the surrounding vicinity. I believe I 
stopped at first stone but it was good enough to be able to store the 
adjacent corners in 19x19 openings.

> In order to get a high "pro prediction" (relevancy) with an "exact 
> matching" pattern-based approach, you need millions of patterns at 
> minumum.
I'd like to echo this point. I didn't have enough, so I had layered databases:
  1. Large radius pro games only
  2. Smaller radius, pro games only
  3. Smaller radius, all games.

Without the third I quickly ran out of moves to play. It played a good 
fuseki though.

> Another thing I found was that small pattern like 4x4 are of limited 
> use. 5x5 is better, but only when you include info on # liberties & stones.
And I'd like to echo this, but my conclusion was liberties is not enough and 
you need tactical search results for every chain and every point in the 
pattern (you mentioned the need for ladder break information in your senseis 
article but that is just the start in my opinion).

When you add liberties to your patterns you reduce the number that match. So 
you need to make them smaller to have enough to get you through the game. In 
my experiments the gain was mostly matched by the loss.

It was at that point I started to run out of time to work on my program, but 
since I restarted a couple of years ago my approach is to get as much 
information as I can from tactical search and life/death analysis: the 
pattern matcher just sits on top and is there to correct mistakes. Give me a 
few more years and I'll tell you if it works. By the way adding tactical 
search slows down pattern training and testing by orders of magnitude. 
Nothing is ever easy in computer go :-).

Katsunari is another program based on using patterns, and the authors 
(Shinichi Sei, Toshiaki Kawashimaha) have written a few papers in English:
  http://homepage1.nifty.com/Ike/katsunari/index_e.html

Darren




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