|
BGonline.org Forums
Rollouts
Posted By: Daniel Murphy In Response To: Rollouts (Stick)
Date: Thursday, 26 July 2012, at 10:55 p.m.
Without knowing the settings and variances in your own rollouts, it's hard to know what "Some of these plays are closer or further apart than [Dmitriy's] rollouts suggest" and "skimping" means. Could you amplify, please?
We know, for example, that if Play A is reported as 0.044 better than Play B with 95% confidence intervals of ± 0.008 for both plays, we shouldn't be surprised if a extended rollout with the same settings reported Play A better by as much as 0.060 or as little as 0.028. Or -- as in Dmitriy's closest result -- if Play A is reported as only 0.004 better with 95% confidence intervals of ± 0.008 for both plays, Play A might easily be 0.020 better, or 0.012 worse, in an extended rollout with the same settings.
Dmitriy's 6 rollouts are all "Moves and cube decisions: 3-ply" with 5,184 trials. Is XG 3-ply play and cube not good enough?
The 95% confidence intervals for the 25 plays Dmitriy rolled out are:
± 0.007 3 plays ± 0.008 13 plays ± 0.009 8 plays ± 0.010 1 play and the differences reported between best and second best play are:
-0.004 -0.021 -0.031 -0.044 -0.045 -0.049 ... which seem like pretty robust results, all in all. One decision is pick'em, one is close but might (or not) be closer, and the other four might be closer (or not) but the best play seems clear enough.
|
BGonline.org Forums is maintained by Stick with WebBBS 5.12.