Re: [Bug-gnubg] Rollout data not displayed in no-gui version

[Jim Segrave]

On Fri 10 Oct 2003 (12:06 +0000), Joern Thyssen wrote:
> On Fri, Oct 10, 2003 at 10:33:39AM -0000, Ian Shaw wrote
> > > Jim Segrave wrote:
> > > 
> > > You are now about to roll out the choice of doubling or not, from the
> > > starting position, with the first roll not being a double. Do the
> > > rollout. The results of the no-double rollout are what you
> > > wanted. They can be saved, printed, extended, what have you. It's just
> > > a shame that you also get a rollout of the illegal double, which is
> > > simply wasted CPU cycles.
> > > 
> > 
> > Thanks, Jim. I'd worked out a similar sequence to get back to the
> > initial position. The only problem left is, as you say, the CPU time. In
> > tests, Analyse...Rollout got through 165 trials in 1 minute, but rolling
> > out the decision only managed 52 trials. I'd expected it to be no more
> > than twice as slow, but it seems like the factor is over 3. Most moves
> > will be the same irrespective of cube position, so I'd have expected
> > caching to reduce the impact of rolling out both cube locations.
> > 
> > I'm looking to do over 60000 trials at 2-ply, so upping it from 6 weeks
> > to 18 is a big deal, nay, heartbreaking.
> > 
> > Is there any way of rolling out take only, or double only? (I'm rolling
> > out post-Crawford 2-away, so I'm interested in the cube being on 2 at
> > the moment.) I can see this being a generally useful option. Many times
> > we are sure of the double, and are only interested in the take decision.
> > Other times, we will be sure it's a take, but wonder about the double.
> 
> "rollout" instead of "rollout =cube" will rollout the current position
> as "on roll", so if you want to rollout the take decision just give the
> opponent the cube and select "rollout".
> 
> AFAIK, this is a function that Snowie doesn't have: snowie can only
> rollout decisions, either cube or chequer play.

The trouble is, it's not associated with a move, so it can't be saved,
exported easily, extended or whatever. What he needs is to be able to
set an initial position, double, return to the double and roll-out
only the no-double portion of the move with rollout as initial. 

RolloutGeneral() doesn't know it's being asked to do a cube decision
rollout, it's simply handed two things to roll-out which happen to be
the two sides of a cube decision[1]. If there was a way to call
RolloutGeneral for a cube decision and set the paramater alternatives
to 1 instead of 2, it would happily only roll out the no-double
portion. For what Ian wants - the various probabilities and the
cubeful/cubeless equities for this side, that would work. The rest of
the output would be problematic, as GeneralCubeDecisionR(), which
makes the call will expect both the no-double and double results to
have been filled in, which they won't have been.

[1] actually it is given a parameter to say whether or not it's a cube
    rollout because it has to use the pre-double cubeinfo structure in
    the rollout code.

-- 
Jim Segrave           address@hidden