Re: [Bug-gnubg] BUG: Rollout statistics wrong - Another Secondary Point

[Jim Segrave]

On Thu 11 Mar 2004 (13:03 -0800), Ned Cross wrote:
> 
> When stopping, then resuming a rollout of several candidate checker plays,
> how can one be sure all the candidates see he same number of trials?

> Scenario:  I am running a long 2-ply rollout, 1296 trials, for 4 candidate
> checker plays.  I need to stop the rollout for whatever reason after only
> 349 trials.  When I hit stop, It's not clear how many of the 4 candidates
> have actually completed the 349th trial.  When I resume the rollout (by
> selecting the candidate checker plays and hitting "rollout" from the
> annotation window), GNUBG opens the rollout window, and after a pause, tells
> me that we are now on trial 350, starting at the first candidate.

The number of trials is saved for each position. When restarting,
gnubg tries to bring every position up to the same number of
rollouts. If it stopped rolling one move out because it was outside
the jsd window, it will still roll that position out for one trial,
then calculate the jsd.

> It seems possible that GNUBG is truncating the trial when "stop" is hit,
> then resuming at the next trial, forgetting about the unfinished business in
> the previous trial.  If so, the rollout results could be skewed.

A particular trial, if interrupted, won't have it's
statistics or its count of trials updated. When you go to resume, the
interrupted trial is redone.

> Also the estimated time remaining is off when resuming a rollout, since GNU
> thinks it has completed 350 trials (the example scenario) in several
> seconds.  This is of course very minor and does not affect the integrity of
> the result.

Yes, it will take a while for the estimate to converge on thereal
value. Elapsed time is not part of the saved information, after all,
you could extend a rollout on a different PC (how's that for a reason
invented after the fact?).

-- 
Jim Segrave           address@hidden