Re: [Bug-gnubg] Extending rollouts

[Jim Segrave]

On Sat 09 Aug 2003 (14:09 -0300), Albert Silver wrote:
> I had a couple of questions regarding extending rollouts.
> 
> 1) If I continue a rollout that was using quasi-random dice, will it
> also continue the proper first roll? Ex: if the first 2 rolls in order
> are 1-1 and 2-1 1-2, and I resume a rollout at trial 38 will it start
> with a 2-1? If not, it should IMO.

It will. It saves the rng seed and the other information needed to
resume exactly where it left off. I tested this earlier by having
gnubg print out the first 6 rolls of every game being rolled out and
comparing an interrupted rollout with a normal one.

I believe (but haven't tested), that this would also work for non
quasi-random rollouts.

> 2) Since I certainly use different settings according to the position,
> and save them in the file, will it know and use the settings that were
> used or will it try to add rollout results using completely different
> settings? Again, if it's not already the case, I think it should use the
> exact same settings as were used. It's almost impossible to put this on
> the user's shoulders since apart from looking into the export results,
> it's impossible to see what the settings used were.

It saves the complete settings (except for the number of trials you
asked for and the various options for stopping a rollout based on
std. or j.s.d. So you can load a match, go to a rolled out analysis,
click on one or more of the moves rolled out, then click the tab for
rollout settings. 

Set a new number of trials and, if appropriate, the stop on JSD or
stop when error is small enough. Don't worry about any other setting.

While it's rolling out, it will 
 save the rollout settings you had before
 install the settings it used for the original rollout (except for the
 trials and stopping conditions)
 roll out the moves
 when it's done, it will put the rollout settings back to where they
 were.

I've had some discussions with Zorba, who would like some better
controls or warnings over this feature, as he often makes changes in
his rollout settings and does usually does not want the rollouts to
extend on their previous settings.

My current thinking is to have an option whether or not it should warn
you. If it's set and you ask to rollout out a move/cube decision which
already has a rollout, the code will compare the current rollout
settings with the ones used before - if they're the same, it will
simply extend the rollout. If they are different, then you get a
dialogue box with some options:

Discard old rollout and use current settings  [ ]
Extend rollout using  previous settings       [ ]
Copy settings from rollout to current         [ ]
Don't give this warning any more              [ ]
Cancel and don't rollout position(s)          [ ]


-- 
Jim Segrave           address@hidden