Miscellaneous: paladin /sk defense skill bug-230

[koros]

Originally Posted by Raev [You must be logged in to view images. Log in or Register.]

Well, you can establish a CI with any number of samples. The problem is shrinking it. I've been gradually putzing around there, and you can see that even with some hundreds of samples the CIs are quite large, assuming I implemented the Wilson Interval correctly.

A glimmering drake [] vs Loin [Various]
Samples: 437
Cooldown: 1.326
Defended: 12.4 - 16.7 (90%, average 14.4, 63 / 437)
Hit Rate: 48.6 - 54.8 (90%, average 51.7, 226 / 437)
Damage: 180 - 380
Per Hit: 233.9
Per Swing: 121.0
Min Hit: 18.0 - 24.9 (90%, average 21.2, 48 / 226)
Interval: 64.0 - 72.0 (90%, average 68.1, 154 / 226)
Max Hit: 8.3 - 13.5 (90%, average 10.6, 24 / 226)

A glimmering drake [] vs Sakuragi [Various]
Samples: 1410
Cooldown: 1.306
Defended: 14.2 - 16.7 (90%, average 15.4, 217 / 1410)
Hit Rate: 34.7 - 38.0 (90%, average 36.3, 512 / 1410)
Damage: 180 - 396
Per Hit: 257.7
Per Swing: 93.6
Min Hit: 22.4 - 27.3 (90%, average 24.8, 127 / 512)
Interval: 70.5 - 75.5 (90%, average 73.0, 374 / 512)
Max Hit: 1.5 - 3.1 (90%, average 2.1, 11 / 512)

Loin is in the early stages of gearing up (Champion's and not much in his other slots), so Sakuragi is probably 100-150 AC (displayed) ahead, yet even with 2000 total samples the 90% CIs (not even the 95% CIs) already overlap for min hits for example.

I think the problem here is this method assumes the distribution to be normal. This isn't the case with EQ. I think the better way to do this is to assign each hit to a 1-20 mapping within the damage interval and use a bootstrapping method. Also since these logs have different max hits, the drakes must be different level. This is going to skew things wildly.

The math behind the ac equation is probably pretty linear, so we can almost certainly use a lower level mob to determine an equation instead of hard to parse drakes.

[Raev]

Originally Posted by koros [You must be logged in to view images. Log in or Register.]

I think the problem here is this method assumes the distribution to be normal. This isn't the case with EQ. I think the better way to do this is to assign each hit to a 1-20 mapping within the damage interval and use a bootstrapping method. Also since these logs have different max hits, the drakes must be different level. This is going to skew things wildly.

The math behind the ac equation is probably pretty linear, so we can almost certainly use a lower level mob to determine an equation instead of hard to parse drakes.

The differing levels are definitely an issue, and I'm probably going to fix that at some point. Sakuragi's parse is actually a combination of multiple fights in an effort to increase sample size (low level glimmers are 180-380, high level are 196-396, and there are some in between I think).

I'm not sure what you mean by a normal distribution; I'm already removing the D20. Min hits in the first are 180, max hits are 380, and interval hits are all other hits. I did kind of mutilate the Wilson Interval by representing it as a binomial rather than a multinomial but my suspicion is that a bootstrap method wouldn't reduce the confidence interval massively. However, I'd be curious to see anything you come up with there.