| |
BGonline.org Forums
Is XG confidence confident?
Posted By: mtuhtan
Date: Thursday, 6 May 2010, at 3:27 p.m.
In some earlier thread someone (sorry I don't remember who exactly) said that he thinks XG's "95% confidence" is not 95%, but somewhat lower.
This is data from one of my rollouts:
equity confidence minimum maximum 1296 trials 0,094 ± 0,023 0,071 0,117 2592 trials 0,118 ± 0,016 0,102 0,134 5k trials 0,131 ± 0,011 0,120 0,142 10k trials 0,134 ± 0,008 0,126 0,142
After 1296 trials, I saw equity and confidence, and almost rejected selected play because of it's low equity (compared to other candidate play).
But, I extended it to 2592 trials to have little better confidence - and now equity is beyond earlier limits for maximum equity!
I thought: OK, confidence is "only" 95%, this is possible (after all 0,118 is very close to 0,117). But now it says maximum equity is 0,134.
Maybe I should extend it even more so equity will drop, closer to first value.After 5k trials, equity is 0,131 (now considerably more than 0,117) and maximum again raised, to 0,142.
After 10k trials, equity is 0,134 (even more than 0,117, and it's same as 2k maximum), with same maximum as 5k - 0,142.
Difference between 10k 0,134 equity and 1k 0,094 equity is difference between close candidate play and (almost) blunder!
Let's say that we can trust 10k rollout.
That means that confidence for 1296 trials should be ±0,048 (100% bigger than showed).
For 2k trials it should be ±0,024 (50% bigger than showed).But if we cannot trust 1k and 2k rollouts, who says that confidence in 10k is correct?
I know there are mathematicians on this forum and I hope to hear their explanation of this situation.
Is this normal behavior for 95% confidence?
Can we trust 1k and 2k rollouts? If not, how about 5k? 10k?
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.