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

[Peter Drake]

On Jul 8, 2006, at 9:40 AM, Don Dailey wrote:

> But unless you find a fuzzy way to do this with patterns,  there is
> always the threshold problem.   Whatever "resolution" you choose you
> will have cases that are on the boundary.    For instance you might  
> have
> 3 cases:  clearly owned by black, neutral, and clearly owned by white.
> But there would be cases near the threshold.  A point might be "almost
> clearly owned by white" for instance.

Yes, but hopefully most cases will not be near the boundary.  We're  
using five levels, with thresholds at -0.6, -0.2, 0.2, and 0.6.

> But I think it's probably a good idea if you can find a good way to
> "generalize" patterns.   If you don't generalize, you have a huge  
> number
> of potential patterns.

Our pattern matcher allows us to generalize to any subset of the five  
ownership bins.  For example, we might match if a point is neutral or  
better, or if it's strong for either player.  (In fact, a decent tiny  
pattern is "don't play on a point that is strongly controlled by  
either player".)

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