RE: [Bug-gnubg] nested mt tasks

[Ian Shaw]

> -----Original Message-----
> From:  Christian Anthon
> Sent: 05 October 2009 22:50
> 
> Hi Jon and the rest of you,
> 
> what is your take on the subject. I would like to implement 
> xgroller functionallity for play, hint, analysis and rollouts. One
> 
> a) call scoremoverollout within asyncfindMove and similar for cubes.
> b) rollouts within rollouts.
> c) rollouts within analysis.
> 
> Probably one would do a) multithreaded while b) and c) are 
> better as single threaded since the parent tasks are already threaded.
> 
> rollouts within rollouts poses its own set of problems 
> becuase of global variables and options.
> 

This option b, rollouts within rollouts, is a strange one. It's not
clear to me how this compares with just doing more trials with a
conventional n-ply rollout. Do you theoretically get more accuracy in
the final result for a given amount of rollout trials.

Obviously, for the same number of trials, you would get a better result
if the rollout player is stronger than the n-ply player, but there would
be a larger variance (I think) and the time per trial would be longer,
too.

I can see that it would be a useful option to have, and some testing
would show how useful it is, but my brain aches just thinking about the
theory of it!

-- Ian