[computer-go] Rotate the dog or the tail?

[Chrilly]

Last year, together with Bruno Bouzy, we published an article in which
we store the 8 orientations of one pattern. In this paper you will also
find several ideas, which I hope you will like.

It can be found at this address, on the website of Bruno Bouzy:
http://www.math-info.univ-paris5.fr/~bouzy/publications/bouzy-chaslot-ci
g05.pdf

Thanks for the link.

I think you underestimate the number of patterns that have to be stored.
In Indigo (Bronze Medal in the Olympiad this year on 19*19), the number
of automatically learnt patterns is about 100,000. And this can still be
increased significantly (cf the paper).

I want to restrict the number of patterns to a few hundred handpicked 
patterns in Suzie.  The Go-expert produces about 10-20 patterns a day. He 
would have to work for months to produce more. This is too much for an "lets 
try if we can make something interesting" project. Mark Boon mentioned in a 
paper that 80 basic patterns are important, additional patterns have only a 
marginal influence on playing strenght.
I think one can not handle 100.000 patterns in a usefull way. The patterns 
and their weight must be consistent. I assume that one sums up a lot of 
white-noise with so many pattern. But maybe has the sum of white-noise some 
usefull information/behaviour. This is  in chess the case. A programm with a 
random eval which searches k-plys plays much better than a programm which 
selects the moves at random. The explanation is, that the eval is an 
indirect measure of mobility. If a side has a lot of moves available, the 
expected value of the maximum is higher.
This has significant consequences. A "normal" eval has a systematic part 
which estimates the true-eval and a random error term. If the tree is 
larger, the expected value of the error-term is higher. The shape of the 
search tree has therefore significant influence on the playing style. E.g. 
if one extends capture moves, the tendency to play capture-moves (or to 
include them in the main-variation) encreases. If one extends pawn-pushed 
which directly attack an opponent piece, the programm starts to like these 
"body-check" moves. This has nothing to do with tactics, but with the 
maximum of the random-term-effect.

Chrilly