[computer-go] Monte Carlo result as a feature for pattern matching

[Peter Drake]

A number of researchers are using Monte Carlo Go as a way of  
estimating the value of a given board position.

(For those of you not familiar with this technique, it involves  
playing random moves to the end of the game and then counting the  
score.  This is repeated many times and an average result taken.)

Pattern matchers regularly look at certain features such as the color  
and edge/off-board status of each point within the pattern.  Some  
pattern matchers look at more sophisticated features such as the  
number of liberties a block has.

We would like to have access to life/death and influence information  
in patterns.  For example, we would like to have patterns that say  
things like "if a weak friendly stone shares a liberty with a strong  
friendly stone, consider connecting them" or "if a point adjacent to  
one of my stones is strongly influenced by me, playing on that point  
is probably not important".

The problem is that open-ended life checkers and influence detectors  
are quite difficult to write.

I propose (and am working on) using Monte Carlo analysis to find the  
expected ownership of each point on the board (say, as a number in  
the range [-1.0, 1.0]) and offering this information as a feature in  
the pattern detector.

Thoughts?

Peter Drake
Assistant Professor of Computer Science
Lewis & Clark College
http://www.lclark.edu/~drake/