Re: [Bug-gnubg] Problem using command files for rollouts.

[Jim Segrave]

On Wed 26 Feb 2003 (08:52 +0100), Nis wrote:
> --On Wednesday, February 26, 2003 00:24 +0100 Jim Segrave 
> <address@hidden> wrote:
> 
> >>I don't know if this is in the wishlist already: "Rollout until
> >>certain".  This should probably be combined with a minimum and maximum
> >>number of  trials.
> >
> >Would this be for a single rollout - where you'd say something like
> >minimum 100, max 2592, stop if STD < 0.01 * abs(equity)?
> >
> >I would think this would be most useful when rolling out a set of
> >moves to decide which is the best move. In this case, some real
> >surgery is required, as you'd then want to stop when say x * JSD <
> >abs(equity of best play - equity of 2nd best play).
> 
> I consider both of these useful. As you write yourself, the second option 
> is the most obviously cool one. What is JSD, by the way?

Joint standard deviation.

 Warning - non statistician speaking:

given a pair of moves with equities e1 and e2, each having stds of s1,
s2, how sure are you that the difference in equities of e1 and e2 is
significant? 

I am told by those who should know that you can calculate 
  jsd = sqrt (s1 ^ 2 + s2 ^ 2)

to get a confidence on the equity difference.
> 
> There is one small complication: What if the top candidate drops in equity 
> when the rollout is extended, causing a previously dropped move to be 
> within range? Should we then resume the rollout of this move also?

My inclination, being lazy woould be to flag a warning that the
results are suspect. Assuming there's a way to extend rollouts (needed
to allow this sort of operation), then the user can extend the
rollouts of the dropped moves to match the moves still in the list
manually.

> >At the moment, you can use the
> >hint or analysis window to get a list of moves, then select a block of
> >those (I don't think you can select non-adjacent ones for a single
> >rollout).
> 
> Ctrl-click to select/deselect individual moves works in the Windows version.


Thanks - I didn't know that

> >After that the current code is very simple minded - roll
> >each selected move out to the limit.
> 
> Yes - and option 1 above would not be very hard to implement either - it 
> "just" needs a new break-condition after each trial.

Yes. When I'm working on the ability to save and resume rollouts, I'll
try to put something in for this.

-- 
Jim Segrave           address@hidden