| |
BGonline.org Forums
more
Posted By: Chuck Bower In Response To: Could this site use a FAQ/guide to posting? (Timothy Chow)
Date: Sunday, 3 April 2011, at 3:38 p.m.
Who should post the rollout result for a position? The primary responsibility rests with the person who posts the problem.
When should the rollout result be posted? Allow 24 hour minimum before posting the rollout result. Maximum of 48 hours should also be a goal.
Is it OK for a person who didn't post the original position to post a rollout? Yes, but make sure to respect the 24 hour minimum. If for some reason the originator doesn't post a rollout within 48 hours, others are encouraged to do so.
Is it OK to post robot evaluations for the problem at hand or rollouts of similar positions? Yes, but make sure to include in the heading (=subject line) "(bot content)". That way others can make their decisions (and potentially post them) before being biased by the bots.
(I'm SURE the following won't generate discussion. :)
What rollout settings should I use?
For XG: 3-ply play, 3-ply cube. (Better but slower: 3-ply play, 4-ply cube.)
For GNU-bg: 2-ply play, 2-ply cube. (Standard "World Class" is good. "Supremo" is better, but slower.)
How long should I roll out a position?
For XG: This bot reports its uncertainty values as 95% confidence intervals (I) with the same units (same scale) as the equity values. For example: low < equity < high, or result +/- I/2. Roll out in increments of 1296 (or multiples of 1296).
a) recommendation 1: for comparing two branches (candidates), make sure 1.5*(larger of the two corresponding confidence half intervals) < difference in equity between the two plays. If not, extend the rollout.
b) recommendation 2: (Stronger than -a-) Make sure the two confidence intervals don't overlap.
c) recommendation 3: (Roughly equivalent to -b-) make sure 2*(larger of the two corresponding confidence half intervals) < difference in equity between the two plays.
for GNU-bg: This bot reports "standard error" (SE), which is (m/l) equivalent to "single standard deviation", also called "sigma". In the report, the appropriate SE is located under the quantity it pertains to, with appropriate units (i.e. scale). Since GNU-bg is considerably slower than XG, 1296 trial minimum takes much longer and may not be feasible. You can start with that value and stop the rollout before it reaches there. 324 trials (= 1296/4) is a decent minimum, but you still should make a goal of reaching statistical significance as follows.
a) recommendation 1: for comparing two branches (candidates), make sure 3*(larger of the two corresponding SE's) < difference in equity between the two plays. If not, extend the rollout.
b) recommendation 2: (stronger than -a-) for comparing two branches (candidates), make sure 4*(larger of the two corresponding SE's) < difference in equity between the two plays.
NOTE 1 (for either bot): If comparing more than two plays, use recommendations above to compare results of the top play (highest finisher in the rollout) to each of the others.
NOTE 2 (for either bot): Sometimes plays are so close in equity that you run out of clock time, or just have better things for your computer to do. In those cases, go ahead and post the rollout but point out to others that you haven't reached statistical significance.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.