Re: [gnugo-devel] Research

[Gunnar Farnebäck]

Doug wrote:
> You might take a look at patterns/extract_fuseki with the statistics code
> I've added recently, which was aimed at that comment. The idea is to
> select a set of games played by GNU Go, collect statistics, and identify
> whole-board patterns which do better or worse than the alternatives. What
> to do then is a separate question, which may depend on things like the
> opponents, the size and significance of the difference in outcome, as well
> as possible reasons for the difference. So far, the only change has been
> to deweight a 3-3 or 3-4 first move on a 9x9 board, but more is no doubt
> possible.

The main limitation of this whole-board approach is of course that
it doesn't generalize at all to novel positions.

I think it would be interesting to try using smaller patterns, maybe
about half-board, maybe smaller, where the rest of the board is
reduced to just a few values, probably related to influence. The
extract_fuseki code can be adapted to extract such patterns from a
collection of games. Matching them in GNU Go while playing is no
problem either. What is an open question is how to determine the value
of moves found through these patterns.

/Gunnar