RE: computer-go: Pattern matching

[Darren Cook <darren@xxxxxxxxxxxxxx>]

>Often several patterns suggest the same move, perhaps 3 to 8.  And a
>symmetrical pattern can suggest the same move twice.

Do you give any extra weight (when move ordering) to a move that was
suggested by more than one pattern?

>I think you need to be a strong go player, or have one to
>suggest patterns.  The strong programs are all written by

Another approach is to learn patterns from professional game records. You
lose the ability to specify the meaning of the move (e.g. David's example
of a running move), but you can make up for this with quantity. Katsunari
(a program by Shinichi Sei and Toshiaki Kawashima) is upto 300,000 patterns
and is apparently getting strong now [1].

My own feeling was that quantity isn't enough, and the program has to try
to understand the meaning of the move in some way when it is learning the
pattern (which as you'd expect is proving hard, and dramatically slowing
down the learning time). But I don't think my experiments ever went beyond
learning about 2000 patterns, so maybe I should consider using a much
bigger training set.

Darren

[1]. From a conversation at GPW'99. If I understood him correctly, Shinichi
also said that the increase in strength had come from improvements in
accuracy (evaluation function?) more than from adding more patterns.