| |
BGonline.org Forums
Fast Rollouts
Posted By: Bob Koca In Response To: Fast Rollouts (neilkaz)
Date: Sunday, 6 December 2009, at 4:31 p.m.
It can be tested. Rollout a position to death to see its true value. Then rollout it out in sets of say 100 games each. For each find the 50% confidence interval and find out what % of the time it does in fact "capture" the true value. I am doing something like that now and will have results in a few days. I don't see why the reported std errors would not be appropriate, the math of how they are made seems fine.
I strongly feel though that the reported joint standard deviation is usually too large as the calculation is based on independence and ignores the advantage gained by using a parallel dice stream.
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.