[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums

Longer Rollout?

Posted By: Daniel Murphy
Date: Friday, 2 November 2012, at 6:38 p.m.

In Response To: Longer Rollout? (Robert Andersson)

There's nothing wrong with rolling out an action until the difference and the confidence interval(s) are all small, but still too large to say with any certainty whether one action is better than the other. But when the difference is small and confidence intervals are large -- as in this rollout so far -- we shouldn't conclude from the rollout even that much.

As a rule of thumb, quadrupling the number of trials should reduce confidence intervals by half. After 429 trials the difference is ±0.001 and the intervals are ±0.013 and ±0.017. To get that ±0.017 down to about ±0.004, the rollout needs to be extended to about 429 * 4 * 4 = 6,864 trials. Even then, the difference between plays after the rollout might still be very small, but if so you'll have a reason to conclude that the cube action is a close decision. You don't have that now -- the rollout after 429 trials does not support a conclusion that the position is a borderline double. It may very well be, but the rollout doesn't say so.

Messages In This Thread

 

Post Response
Your Name:
Your E-Mail Address:
Subject:
Message:

If necessary, enter your password below:

Password:
 

 

[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums is maintained by Stick with WebBBS 5.12.