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

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

From: "Vincent Diepeveen" <diep@xxxxxxxxxxxxxxxxx>
Subject: Re: [computer-go] Pattern matching - example play



> Yet your evaluation function needs to understand what type of eye it is
and
> just the word 'eye' won't do in your evaluation function. You need clear
> distinctions between what type of eye it is.


Yes. But my idea was to have an almost infinite number of "eyes" and let the
harvester/learner assign an almost infinite number of values to them.

I don't see why it should be much more complicated than harvesting millions
of patterns and assigning by an analysis method values to them. I started
with pattern because it was the most basic and the easiest.

With eye shapes I would go over half a million games and look at all chains
and the groups they form (for ex. by their connectivity status that has been
established by automatic learning of connectivity, which works quite well,
I've already done it).

Then I would devise a method of translating the eye shapes into a hash value
(the actual shape, like a pattern).

Then I would read many publications on anything that could be important with
eye shapes.

Which color is to move is important, so that should be hased in, etc.

Then of course what needs to be known is the status of the group or chain,
at the end of the game. Did it succeed in making an actual eye at the
location of the potential eye? How often did that succeed for each potential
eye pattern?

If not, why not? Was it due to the fact that this potential eye "stunk" :)
or was it due to the fact that more pressing business was needed to be
tended to, like defending some larger chain etc.

Even without that refined knowledge, simple statistical analysis of how
often a potential eye becomes a real eye is a real "eye-opener" :)

Then you'll end up with say 8 million possible eye shapes and statistical
analysis based on half a million games on the likelihood that each shape
will become a real eye.

And people really want to tell me that THEIR knowledge on eye shapes is more
varied than 8 million shapes and more reliable that statistical analysis on
half a million games?

Sorry but I really believe that when you combine 8 million patterns with 8
million eye shapes and 8 million connectivity patterns and 8 million
territory patterns etc. etc. and you use that stuff in a good search
algorithm, that it will blow away anything ever made.

There is no way in hell that some manually entered eyeshape lib will
outperform millions of thoroughly analized eyeshapes.

I think I have already shown that an extracted/learned pattern library is
able to predict 46% of any pro move in a random pro game on average, so why
would a learned eyeshape lib perform worse than a manyally entered one?

Imagine the synergetic power of every learned module.

The pattern module already performs at 46% but makes bad moves because it
has no clue about eyes and that is a particularly weak point. So to the
rescue an eyeshape expert system that overrules the pattern module. You can
LEARN when which pattern is more important, the "pshape" pattern or the
"eye" pattern by simply letting the modules go side by side, predicting
moves and adjusting their pattern's values with the moves that are actually
played.




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