Multi-hit WS accuracy testing

[CDF]

The lack-of-confidence statement seems to be based on an explicit comparison of the WS hit frequencies alone, but there is no particularly good reason to believe that there was any difference in hit rate between secondary WS hits without gorget and auto-attack hits. It doesn't make sense to ignore the 10,000 melee hit attempts as though that information weren't available (this is all assuming the same-level clots were used, if there was more than one clot).

There is no reason either to believe there is any difference between off-hand hits from dual wielding and secondary WS hits in terms of hit rate. This is not to say that there was sufficient sample size to "prove" no difference, but that wasn't the goal of the data collection and it's kind of "wasteful" to treat the data conservatively.

[Kirschy]

Yea, I mostly agree with the fact the off hand hit, and extra hit are the same hit% as the normal hits. It just isn't enough to convince me that the gorget adds +10 accuracy to all multihits. It's enough evidence for most, and I can understand why. But I'm still personalli skeptical, and I'd like stronger evidence and larger sample sizes.

I'll stop crapping up this thread /w my pessimism, at least until I have some data or evidence to provide one way or the other~

[Max™]

If you're admitting to using a PS2 for data collection than I wouldn't advise anybody to take the results seriously, especially considering how inappropriately jaded you are against other forum posters (I'm not taking sides either, it's a ridiculous and childish situation to begin with.)

I'm only jaded against Starr, I respect Thorny's mathematical prowess, but I also understand enough about math to do the piddily stuff involved with FFXI formulas myself.

I recently straightened all my shit up in my room, and my notebook is one among many in a drawer now, it's been something like a year and a half since I did the testing.


I remember sharing it with Ruke though, posting the TP returns I got, the damages by occurance rate as it was a static target (done over several hours nonstop on a garlaige citadel [S] fort).

Changing nothing but my neck piece from empty > snow gorget, I went 80% Acc to 85% Acc over two sets of 200 WSes without and 200 WSes with gorget.

I did another set without brutal and with claymore grip to reduce my DA rate to minimal as Ruke mentioned it may have skewed the results.

I had a /echo <tp> macro after I WSed, and changed no gear to maintain the same hitrate as my meleeing.

I made a tally mark in my notebook next to each of my returns: 1%, 2%, 3%, 4%, 16%, 17%, 18%, 19%, 20%, 21% I think.

I noted but did not count the returns that landed the first hit in the 200 WSes with/without, though the TP returns including the first hits did line up with the same acc boost, but I had to over a thousand WSes to get the 400 with first hit misses.

Know what a parser does? It records the data, same thing a bored geek with a notebook does.

[pchan]

The margin of error depends on the hit rate you are measuring. If your rate is approxiamtely 50%, with 200 sample size you get ~ +/- 7% error, and if you have 90% than it's +/- 4%. So if you go from 50% to 55% you cannot statistically conclude +10 acc, it can be anything between 0 and 15 pretty much..

[Starr]

I'm only jaded against Starr

Because I talk to you like an idiot, because you are an idiot? Tough world we live in.

[Thorny]

Max's notebooks may sound ridiculous, but it's still pretty funny to hear Starr calling him an idiot.

[Starr]

Explain.

[Anomander]

I'm inclined to agree with Max here, my own parses seem to agree (mostly based on rana). Guillotine definitely was hit with an acc penalty a few updates back, as someone mentioned, and was a poor choice for the OP's test.

[Starr]

Could you post these parses?

[CDF]

The accuracy penalty for Guillotine and Penta Thrust was allegedly removed in the Nov 2007 update after being "adjusted" (downward?) in Sept 2007. Actually read some posts before thinking that the gorget effect was even tested.

[Anomander]

Read the entire thread, don't remember it being mentioned that Guillotine's acc nerf was removed or seeing it in the patch notes (I do recall it for penta). Going to go look into it now, if I'm wrong then I retract my previous statement and say that the results were quite strange.

[Nameless]

The accuracy penalty on guillotine was reduced not removed as far as I can remember. And I think the penalty started at the same time as the 2hander update or the patch after that, and then it was reduced in the very next update. Check those patch notes and you should see them.

[CDF]

The update was after the major version update (Nov 20):

FINAL FANTASY XI Update ( Nov. 28 )

- The penalty to Accuracy for the weapon skills "Guillotine" and "Penta Thrust" has been removed.

To the extent that accuracy modifications on multi-hit weapon skills are not very well understood (whether accuracy increases with TP, whether there is any bonus at 100 TP for the relevant weapon skills, or even any penalties), Guillotine was as good a candidate as any to explore this, whether as a baseline for comparison or as an example of one that may still have some penalty, as it appears.

[Knyghtmayre]

I'm more interested as to how TP per hit on every trailing hit after the first is calculated? 1TP per hit and it's affected by StoreTP? But how, exactly?

[Thunder]

I'm more interested as to how TP per hit on every trailing hit after the first is calculated? 1TP per hit and it's affected by StoreTP? But how, exactly?

If you have +40 stp during WS wouldn't it just sort of be (1*1.4)=1.4 per extra hit? That's how I worked it out when doing hypothetical drk 5hit builds.

Not sure how to test that.. work out how much store tp+ you need on sam for the initial rana hit + the two extra to equal a whole number and then try it again but this time with -1STP?

Or create a 5hit build on say, drk, that requires guillo to hit all three times and try that out a lot. Or Rana with +50 STP until the first hit misses and next two connect and see if you get 3tp return. Then to make sure try it with +49TP and see if you get 2tp return.

[Jojojo]

For the people claiming the sample size wasn't high enough:

Spoiler: show

http://i15.photobucket.com/albums/a3...guillotine.jpg

The probability of getting 129 or fewer hits out of 300 with a 50% hit rate is less than one percent. There's undoubtedly a penalty here.

If people are saying you need more testing in order to pin down the value, well, no shit. You're gonna need a sample size of over 4300 in order to be sure you are within even 5% of it, up or down. That's still a fucking huge range. It is impractical to try to find the exact probability for anything in this game.

[Jojojo]

The margin of error depends on the hit rate you are measuring. If your rate is approxiamtely 50%, with 200 sample size you get ~ +/- 7% error, and if you have 90% than it's +/- 4%. So if you go from 50% to 55% you cannot statistically conclude +10 acc, it can be anything between 0 and 15 pretty much..

p(1-p) is maximized at p=.5, so it's generally okay to assume .5 when dealing with the variance of the sample mean. Even if p is as low as 30% or high as 70%, p(1-p) is still .21, as opposed to .25. Not a big difference. So if you wanna simplify your calculations just use p = .5 ^^

I apologize to everyone else for having to read that.

[Abulurd]

I'm probably doing something wrong and/or useless, but the range for where 129/300 lands at one standard deviation or less from the midpoint is between a ~40.3% and a ~46% hit rate. (I'm leaning towards useless, personally.)

[Byrthnoth]

Floor(Number of hits after the first * STP)

So with 23 STP using Pentathrust, you have:
Floor(4*1.23) = Floor(4.92) = 4.9 TP return, plus the TP from the initial WS hit.
If it was floored individually, you'd only expect a 4.8TP return from the additional hits.



Anyway, the OP's testing was pretty interesting, and confirms what melees have anecdotally suspected for a long time. Large Acc boost on the first hit of every WS, no bonus for the later hits.

I just used a one-tailed student T-test, null population was assumed to be 95% and 50% respectively:
There's a 13.5% chance that the 4% difference between the 91% (measured) and 95% (theoretical) hit rates on the first hit of Guillotine resulted from random variation. So there's an 86.5% chance that there is an accuracy penalty.
There's a 4.3% chance that the 7% difference between the 43% (measured) and 50% (theoretical) hit rates on the latter hits of Guillotine resulted from random variation. So there's a 95.7% chance that there is an accuracy penalty.

The weirdest part about this is that it looks oddly plausible that whatever Acc penalty might be on Guillotine affects the first hit too, or possibly that the first hit Acc bonus just doesn't overcome it enough to cap hit rate. It would have to be somewhere around +90 to cap hit rate without the accuracy penalty, so +100 seems a logical place to start.

[pchan]

p(1-p) is maximized at p=.5, so it's generally okay to assume .5 when dealing with the variance of the sample mean. Even if p is as low as 30% or high as 70%, p(1-p) is still .21, as opposed to .25. Not a big difference. So if you wanna simplify your calculations just use p = .5 ^^

I apologize to everyone else for having to read that.

NO, it's the opposite. The error is the biggest when the variance is the highest. The error is :

1.95* sqrt( p*(1-p) ) / sqrt(N)

N being sample size and p the observed rate.