On Jan 23, 2008 7:39 PM, Jason House <[EMAIL PROTECTED]> wrote:
> On Wed, 2008-01-23 at 18:57 -0500, Eric Boesch wrote:
> > I am curious if any of those of you who have heavy-playout programs
> > would find a benefit from the following modification:
> >
> > >   exp_param = sqrt(0.2); // sqrt(2) times the original parameter value.
> > >   uct = exp_param * sqrt( log(sum of all children playout)
> > >                           * (child-win-rate-2) /
> > >                         (number of child playout) );
> > >   uct_value = (child winning rate) + uct;
> >
> > where child-win-rate-2 is defined as
> >
> > (#wins + 1) / (#wins + #losses + 2)
>
> I'm surprised to see that this works as listed, because the math looks
> all wrong to me...

Argh. I have to retract the claim that this helps. I didn't optimize
the libego parameters correctly before I tested it. Sorry about that
-- I thought I did. There's a lot more I could add, but I thought I'd
get that out there before anyone wasted (probably) any more time on my
error.

By the way, does anybody know of any nifty tools or heuristics for
efficient probabilistic multi-parameter optimization? In other words,
like multi-dimensional optimization, except instead of your function
returning a deterministic value, it returns the result of a Bernoulli
trial, and the heuristic uses those trial results to converge as
rapidly as possible to parameter values that roughly maximize the
success probability. The obvious approach is to cycle through all
dimensions in sequence, treating it as a one-dimensional optimization
problem -- though the best way to optimize in one dimension isn't
obvious to me either -- but just as with the deterministic version of
optimization, I assume it's possible to do better than that. It might
be fun problem to play with, but if good tools already exist then it
wouldn't be very productive for me to waste time reinventing the
broken wheel.
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/

Reply via email to