[Pachi] A little bit lost in the code
jm.arbona at iecb.u-bordeaux.fr
Thu Jan 5 09:40:18 CET 2012
So for the question 2),
Does it mean than when you are on an adversary move you
apply the algorithm on 1-rewards to minimise its regrets?
And even if you explore the branch that make you loose,
you will have a good estimate of the score for this branch, but
the mean rewards of the root node will still be good.
Thanks for your time :-)
2012/1/3 Petr Baudis <pasky at ucw.cz>
> On Tue, Jan 03, 2012 at 02:45:09PM +0100, Jean-m. a. wrote:
> > the HACKING file helped a lot.
> > (I am rosted in c and had trouble to understand the ucb1.c code because
> > the code in generic.h that I didn't found...)
> > I have some questions:
> > 1) to avoid that different processes explore the same branch you had a
> > virtual loss at the beginning of the branch.
> > But it means than only one thread at a time can explore a branch
> > characterised by one starting move.
> > So even if you have 300 thread, more or less each thread is going to
> > explore only one starting move?
> The virtual loss does not have so strong effect. At the very
> beginning, this may be the case, but successful moves near the root
> will quickly attract many/most/all threads despite the virtual loss.
> > 2) A question maybe stupid:
> > If given one starting move allmost all the move of the adversary let you
> > win and only one or two are sure to make you loose,
> > the average winrate must be good. And furthermore you are going to
> > more the subbranch that let you win. How is this effect taken into
> > account?
> It is a minimax tree. If one or two branches of the opponent are sure
> to make you lose, these will be explored the most.
> Petr "Pasky" Baudis
> The goal of Computer Science is to build something that will
> last at least until we've finished building it.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Pachi