Le 26/11/2009 à 10:08, Vlad Dumitrescu a écrit :
> 
> On Thu, Nov 26, 2009 at 00:43, Darren Cook <[email protected]> wrote:
> > When I read this it reminded me of experiments I tried before to pass
> > more than one piece of information up from the leaf nodes of a (min-max)
> > tree. E.g. a territory estimate and an influence estimate. I gave up as
> > it got too complex to handle incomparable nodes (e.g. move A gets more
> > territory, less influence). I remember having a really good reason to
> > want to delay reducing multiple features to a single number , but it is
> > all a bit fuzzy now.
> >
> > Does this type of search have a name, and any associated research?
> 
> It feels like something that should be covered by some mathematical
> area, but I spent half of last night searching in vain. It's possible
> that I just didn't understand what those papers said.
> 
> One issue is that when the value at a node includes other features
> than win/loss information, then the game is no longer zero-sum and
> then maximin should be used instead of minimax. Also the linearization
> of these feature's values should be done differently for each player
> (even if we might choose to play a little recklessly ourselves, for
> the opponent we should probably use the safest line of play).

Maybe have a look at signal processing, using higher-orders statistics ?
 mean 
 std-deviation = order 2 (or 1 ?)
 ...

 win by 10 with std = 100 seems much less secure than win by 5 with std=1
 but maybe this is included in modern bots (i stopped at naive MC + AMAF)

It's far in my memory, so i can't tell you much more
Alain
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/

Reply via email to