Re: [Bug-gnubg] Feature request: Rollout frequency based on std. errors

[Christopher Yep]

Some additional notes:

1. To avoid the slight chance of confusion, I actually should have used 
"play" instead of "move" in my last e-mail.  The original feature request 
was to apply this algorithm for one given position/roll, though ultimately 
it would be even better if it could be applied across multiple positions 
(using CMark, see #5 below).

2. Obviously the algorithm only needs to be applied when the user selects 
two or more plays to rollout.

3. The algorithm could also be applied to cube decisions (gnubg only looks 
at two plays in this case: no-double and double/take).

4. If a user wishes to rollout a move (selecting all reasonable candidate 
plays) until one play is "best" (to some defined level of statistical 
significance and ignoring all systematic errors that gnubg may make during 
the rollouts), this algorithm will almost always reduce the amount of 
rollout time needed.

5. Ultimately, this algorithm could be applied across multiple moves 
(positions): e.g. a user could CMark a large number of candidate plays 
across multiple moves (throughout the entire match).  He could then, for 
example, instruct gnubg to rollout all candidate plays (throughout the 
entire match) until the standard error of every candidate play is less than 
.02, rolling out each candidate play a minimum of 36 times.

Chris