Re: [gnugo-devel] machine learning in gnugo

[Gunnar Farneback]

Thomas Nelson wrote:
> Hi, my name's Tom, and I'm a computer science undergraduate at UT
> Austin. I'm doing a research project on machine learning in go.  My
> plan is to choose ~10 patterns in the gnugo patterns database, and
> optimize the shape() values for the patterns using a kind of gradient
> descent learning. My question is, which patterns should I choose?
> Ideally, I want to choose the 10 that will have the biggest effect on
> gnugo's performance.  They should be used pretty often, and preferably
> early in the game.  My current plan was just to play a lot of games,
> and use
> gnugo -l [filename] -L [move number] -t -a -w
> on each move to see which patterns get selected most often.  But maybe
> someone knows a particular group or type of pattern that needs more
> attention, or a more efficient way to search for important patterns?

I'm doubtful there are any patterns which are really suitable for your
plan. I think it would be more interesting to do the experiment with
some selection of coefficients more centrally involved in the move
valuation, e.g. the ones in connection_value() in value_moves.c, in
crude_dragon_weakness() in dragon.c, and/or in
compute_dragon_weakness_value() in dragon.c.

/Gunnar