A Call to Arms - MoP Mechanics Testing

[benebarba]

quick check of build 15752 (current) shows that holy wrath still appears to reset the swing timer.

Haven't been able to snag another windsong enchant the few times I've logged in - either absurd price (last one was over 75K) or lack of auction.

[theckhd]

Not a test, per se, but a quote that will save us from doing a lot of annoying testing:

Crit chances of players against mobs that are higher level than you are reduced by 1% per level difference, in Mists.

[tlitp]

Armor and physical damage reduction :

• Player level - 90
• Talents/Glyphs - irrelevant
• Gear - Prot/Ret, with high Str/AP; you should have a few spare parts available, as you'll need to shift the total AP (e.g. base total : 10k, and then {10.5k 11k 11.5k 12k})
• Weapon - Club (must be exactly this one), purchasable from vendors
• Buffs/Debuffs - ideally none
• Target - level 93 dummy (ideally less crowded)
• Attack sequence - only autoattacks, from the back
• Goal - equip the base AP gear (note the value somewhere), attack the dummy for 1 min; add/replace one piece (changing the AP, noting the new value), attack for another 1 min; then repeat the last step three more times, to have five data sets

Base stats/attributes : level 90, different races. Being fully naked, report primary stats (Str/Agi/Int/Stam), total AP and physical/spell crit chance (character sheet), total HP and MP.

[theckhd]

Another potentially useful quote:

The two links you mention are old, but the ideas are accurate. White attacks are still ‘one roll’ (Miss Dodge Parry Block Glance Crit Hit), and yellow attacks are still ‘two roll’ (Miss Dodge Parry Block, Crit). The only change to that is that Block against players is an additional separate roll now. Glances can still only occur on white attacks. I believe our recent blog post about this stuff included all off the relevant chances for these, but if you have more specific questions, feel free to ask.

[Klaudandus]

I need data regarding Str->Parry conversions and dodge/parry DR equations. All of that can be had by equipping different amounts of gear and reporting the dodge and parry tooltips (i.e. "X rating gives Y% dodge (before DR)" as well as reporting your overall dodge chance and Strength value).

lvl 90 belf

Str - 4914
Dodge chance of 6.03%
Dodge of 2896 adds 2.80%
Parry Chance - 10.19%
Parry of 1460 adds 1.41%

Str - 5308
Dodge chance of 6.21%
Dodge of 3077 adds 2.97%
Parry chance of 10.73%
Parry of 1559 adds 1.51%

Str - 5483
Dodge chance of 6.54%
Dodge of 3416 adds 3.30%
Parry chance of 10.68%
Parry of 1318 adds 1.27%

Str - 5956
Dodge Chance of 5.45%
Dodge of 2322 adds 2.24%
Parry Chance of 11.19%
Parry of 1305 adds 1.26%

Str - 6123
Dodge chance of 6.26%
Dodge of 3128 adds 3.02%
Parry chance of 11.50%
Parry of 1429 adds 1.38%

Str - 6371
Dodge chance of 5.80%
Dodge of 2668 adds 2.85%
Parry Chance of 11.67
Parry of 1326 adds 1.28%


Str - 6589
Dodge Chance - 6.60%
Dodge of 3476 adds 3.36%
Parry Chance - 11.40%
Parry of 825 adds 0.80%

Str - 6751
Ddoge chance - 6.43%
Ddoge of 3306 adds 3.19%
Parry Chance - 11.58%
Parry of 823 adds 0.80%

[theckhd]

<edit> I should note that I believe there's a typo in your data. The dodge from data set #6 should be 2.58%, not 2.85%. It's the only outlier, and it can't be reconciled with the rest in any way. If I calculate the results without that data point, I get a result that's consistent with 2668 dodge rating giving 2.58% dodge pre-DR.

Based on the rating to percentage conversions, the coefficient for both dodge and parry seems to be the same. The dodge data puts the range between 1034.7 and 1036.0, the parry data puts it between 103.8 and 105.2. Together, we can narrow this range to 1034.7-105.2, making 1035.0 a pretty reasonable estimate (given the accuracy limitations inherent in rounded data). I came about this through simple unit testing, but curve fitting gives a similar result:

General model:
f(x) = x/(1035+a)
Coefficients (with 95% confidence bounds):
a = 0.2393 (-0.4219, 0.9005)

Goodness of fit:
SSE: 0.0001106
R-square: 1
Adjusted R-square: 1
RMSE: 0.002715

The rounding is really the limitation on the accuracy, I think. So we'll need more data to be able to pin it down to better than +/- 0.5.

[theckhd]

I'm having trouble curve-fitting the dodge data. There are a few peculiarities. For example, with 2.8% pre-DR dodge and 3% base dodge, you'd expect less than 5.8% total dodge after DR. Yet we have 6.03%.

It's almost as if they removed diminishing returns completely. I can fit the data pretty well with

Code: Select all

total_dodge = (predr_dodge_from_rating+3)/0.9623

I can also fit it with

Code: Select all

1/total_dodge = 1/C + k/(3+predr_dodge_from_rating)

which is the old formula. k comes out to 0.9517+/-0.01, which is reasonable, but C is 562.1+/-517, which I don't trust (for obvious reasons). C is the cap on avoidance, and is currently around 65ish. That falls within the confidence bounds, but it's not at all confidence-inspiring to have such a large range.

I was hoping to isolate dodge individually to figure out the DR equation, and then use those results to figure out the STR->Parry ratio, but if I'm going to do that I'll need more data at much smaller dodge rating values (i.e. several data points between 100-2000 rating). Or much larger values, for that matter, but that may be harder to do.

[Klaudandus]

so, what do you need me to do? i can sit next to a reforger and try to get as many dodge and parry ratings... would that help?

lvl 90 belf - no buffs
Dodge Chance 3.91%
Dodge of 833 adds 0.80%

Dodge chance 4.25%
Dodge of 1162 adds 1.12%

Dodge chance 4.56%
Dodge of 1457 adds 1.41%

Dodge chance of 4.85%
Dodge of 1738 adds 1.68%

Dodge chance of 5.04%
Dodge of 1924 adds 1.86%

Dodge chance of 5.28%
Dodge of 2155 adds 2.08%

Dodge chance of 5.39%
Dodge of 2261 adds 2.18%

Dodge chance of 5.68%
Dodge of 2553 adds 2.47%

Dodge chance of 5.98%
Dodge of 2853 adds 2.76%

Dodge chance of 6.31%
Dodge of 3184 sffd 3.08%

Dodge chance of 6.58%
Dodge of 3461 adds 3.34%

Dodge chance of 6.86%
Dodge of 3747 adds 3.62%

Dodge chance of 7.07%
Dodge of 3965 adds 3.83%

--
Str - 6430 (will stay the same for the rest of this list)
Parry chance of 10.96%
Parry of 569 adds 0.55%

Parry chance of 11.24%
Parry of 843 adds 0.81%

Parry chance of 11.53%
Parry of 1126 adds 1.09%

Parry chance of 11.79%
Parry of 1380 adds 1.33%

Parry chance of 12.14%
Parry of 1730 adds 1.67%

Parry chance of 12.41%
parry of 1997 adds 1.93%

Parry chance of 12.76%
Parry of 2349 adds 2.27%

Parry chance of 13.07%
Parry of 2658 adds 2.57%

Parry chance of 13.26%
Parry of 2844 adds 2.75%

---
That's as much as I can squeeze out of Parry without changing my str at any point.

[theckhd]

Realized I was making a dumb mistake - I was putting base dodge into the DR formula, when traditionally base dodge isn't affected by DR. That makes a world of difference in the results. Combining the earlier data set with the one that Klaud just provided, here's what I get:

Code: Select all

General model:
     f(x) = 3+1/(1/C+k/(x))
Coefficients (with 95% confidence bounds):
       C =       61.84  (57.08, 66.59)
       k =      0.8791  (0.8753, 0.8829)

Goodness of fit:
  SSE: 0.0005156
  R-square: 1
  Adjusted R-square: 1
  RMSE: 0.005209

If I fix C at it's cataclysm value (near the upper end of the confidence bound on C), I get:

Code: Select all

General model:
     f(x) = 3+1/(1/65.631440+k/(x))
Coefficients (with 95% confidence bounds):
       k =      0.8819  (0.8811, 0.8827)

Goodness of fit:
  SSE: 0.0005827
  R-square: 1
  Adjusted R-square: 1
  RMSE: 0.005398

At this point, it's not completely conclusive - we have two free parameters, and the data doesn't deviate enough from linearity over this range to feel too confident about either, but if I had to guess I'd say they kept the old caps and simply changed the constant k. Note that since k<1, this means that you actually gain avoidance from the DR equation at low levels (like we are here). It isn't until you stack an appreciable amount (probably 10-15%) that it starts to fall behind the pre-DR values. In Cataclysm, k was about 0.965, so this effect was still present at very low dodge/parry (<2%), but decreasing k will make it extend up to higher values.

Working on the strength/parry data now. I was fooling around with a surface fit of the earlier parry data, and the indication is that the parry gained from STR isn't affected by DR. But now that I have some new data to play with, I should be able to tell for sure.

[Klaudandus]

I'll try to get more gear, I'm barely 422 ilvl -- that should help me get more dodge/parry rating

[tlitp]

Base stats/attributes : level 90, different races. Being fully naked, report primary stats (Str/Agi/Int/Stam), total AP and physical/spell crit chance (character sheet), total HP and MP.

From SimC's guts :

Code: Select all

Str   Agi   Sta   Int   Spi
178   105   169   114   123  (human)
183   101   170   113   122  (dwarf)
179   102   169   114   125  (draenei)
175   107   169   117   121  (blood elf)
183   101   170   110   125  (tauren)

base hp : 146663
base mp : 60000

Conversion factors, from SimC's guts :

Code: Select all

agi->ph.crit
1259.51806640

int->sp.crit
2533.66357421

dodge, parry
1035

block
345

phys/spell hit, exp
400

phys/spell crit, mast
700

phys/spell haste
500

The above are DBC data from b15781.



EDIT : gief dataz foar amrorz NAO.

[theckhd]

Parry/Str data is being... confusing. With a curve fit to just the most recent data set (fixed STR), I can get a very good fit from either model:

Code: Select all

General model:
     f(x) = 3+s*6430+1/(1/65.631440+k/(x))
Coefficients (with 95% confidence bounds):
       k =      0.9079  (0.9025, 0.9132)
       s =    0.001144  (0.001142, 0.001146)

Goodness of fit:
  SSE: 0.0002031
  R-square: 1
  Adjusted R-square: 1
  RMSE: 0.005387

Code: Select all

General model:
     f(x) = 3+1/(1/65.631440+k/(x+s*6430))
Coefficients (with 95% confidence bounds):
       k =      0.7108  (0.7061, 0.7154)
       s =    0.000915  (0.0009074, 0.0009227)

Goodness of fit:
  SSE: 0.0002806
  R-square: 0.9999
  Adjusted R-square: 0.9999
  RMSE: 0.006331

However, the extreme difference in s suggests that we should be able to tell the difference between these by simply varying strength. So I took the entire data set and used the surface fit tool. That gives me these situations:

Code: Select all

General model:
     f(x,y) = 3+s*y+1/(1/C+k/(x))
Coefficients (with 95% confidence bounds):
       C =        30.5  (19.26, 41.74)
       k =      0.8448  (0.8034, 0.8862)
       s =    0.001136  (0.001129, 0.001142)

Goodness of fit:
  SSE: 0.003198
  R-square: 0.9997
  Adjusted R-square: 0.9997
  RMSE: 0.01511


Code: Select all

General model:
     f(x,y) = 3+1/(1/C+k/(x+s*y))
Coefficients (with 95% confidence bounds):
       C =  4.404e+004  (-9.616e+006, 9.705e+006)
       k =       1.016  (0.9207, 1.111)
       s =    0.001184  (0.0011, 0.001268)

Goodness of fit:
  SSE: 0.03785
  R-square: 0.9967
  Adjusted R-square: 0.9963
  RMSE: 0.052


If I fix C in the last two versions, I get:

Code: Select all

General model:
     f(x,y) = 3+s*y+1/(1/65.631440+k/(x))
Coefficients (with 95% confidence bounds):
       k =      0.9024  (0.8882, 0.9166)
       s =    0.001144  (0.001139, 0.001148)

Goodness of fit:
  SSE: 0.005202
  R-square: 0.9995
  Adjusted R-square: 0.9995
  RMSE: 0.01862


Code: Select all

General model:
     f(x,y) = 3+1/(1/65.631440+k/(x+s*y))
Coefficients (with 95% confidence bounds):
       k =       0.726  (0.6867, 0.7653)
       s =   0.0009355  (0.0008722, 0.0009987)

Goodness of fit:
  SSE: 0.06458
  R-square: 0.9944
  Adjusted R-square: 0.994
  RMSE: 0.06561


Neither of the versions with strength-based parry affected by DR fit very well. One easy way to check would be to have a few data points at exactly the same parry rating but different Strength values (i.e., full gear, then take off pieces of gear that don't have parry rating on them one by one). That would probably give me a clear strength scaling to work off of. This could be done in ret gear, even, to give the largest STR range possible.

[Klaudandus]

I'll see what I can do about that -- another alternative is to get some pieces of spirit plate gear and reforge into parry.

Gimme a few and I'll do just that, need to find a fricking vendor first AND get more bagspace >=/

Edit: Actually, give me a while -- getting a really bad migraine...

[theckhd]

Conversion factors, from SimC's guts :

Code: Select all

block
345


The above are DBC data from b15781.



EDIT : gief dataz foar amrorz NAO.

Updating the code with these now, but I don't understand the block entry. It looks like mastery rating to mastery is 700, but we get 1% block per mastery unless something's changed. Is the "block" entry for block rating, in case it's still floating around on low-level gear?

Also, if anyone's level 90 on beta and can get the armor data set that tlitp wants, please do so. You wouldn't like him when he's angry.

[Klaudandus]

working on tiltp's logs

Edit: DONE

Working on squeezing varying amounts of parry without changing my STR