Re: [Bug-gnubg] Cmarked rollouts some moves stopping short of set trials

[Philippe Michel]

On Sat, 16 Jan 2010, Christian Anthon wrote:

> As usual an example makes the problem easier to track...

> On Sat, Jan 16, 2010 at 1:47 PM, Michael Depreli
> <address@hidden> wrote:
> > Since updating to build 20091230 I'm having problems with cmarked rollouts
> > stopping
> > short of the set trials for some moves.
> > I'm using 1296 trials minimum 720 if jsd 2.33 is reached.
> >
> > I'm seeing that one move has reached 1296 but the others seem to be stopping
> > at anything between 1295 and 1285 trials typically.  JSD has NOT been
> > reached.

With a setup similar to Michael's (set rollout trials 1296, set rollout 
jsd minimumgames 324), I get the same kind of behaviour all the time. 
Something like :

        1296 games, Mersenne Twister dice gen. with seed 845589805 and 
quasi-random dice
        1291 games, Mersenne Twister dice gen. with seed 845589805 and 
quasi-random dice
        1292 games, Mersenne Twister dice gen. with seed 845589805 and 
quasi-random dice

for three candidate moves, and sometimes it even gets to a little 
below 1290

It doesn't particularly shocks me, although it would be nice if the 
rollouts were stopped at exactly 1296.

I think I read a comment somewhere (maybe in this list or in the code) 
that with multithreading and the current way to stop rollouts this was to 
be expected. Since most recent changes in rollout code are yours, 
Christian, it may even have been you who would have written that.

Maybe these discrepancies are larger than you thought. Maybe they are 
smaller with the more common dual core systems but can get larger with 4 
or more cores ? Above example is with 4 threads.

Best regards,

Philippe Michel