[computer-go] Heuristics for MC/UCT with all-or-nothing payouts

[Peter Drake]

On Jun 10, 2007, at 5:29 PM, Eric Boesch wrote:

> As for experiencing limited improvements by tweaking the formulas, it
> would hardly be surprising if there are limits to how much an MC/UCT
> (or BAST or whatever) type program can be improved by just swapping
> out one formula and replacing it with another one using the same
> variables, since the formulas are pretty good to begin with, and the
> more glaring weaknesses of vanilla MC/UCT, which the strongest MC
> programs already address in other ways (though I think there is still
> room for improvement in their search localization in particular, not
> that I have any clear notion of how the improvement could be achieved
> without jeopardizing the programs' existing strengths), can't be fixed
> by just tweaking formulas.

I agree -- the big gains will come from search localization, breaking  
the game into semi-independent subproblems, or goal-oriented search.  
(Maybe these are three names for the same thing.) Of course, I don't  
know how to do it yet...

Peter Drake
http://www.lclark.edu/~drake/