A DPS Simulator 2 - SimFF

[Viridiana]

Thing is, the first set doesn't seem that off if you ignore the 150 potency point. Example: 1.571875 damage per potency would lead to:

Agreed, but we both know we can't just ignore data all willy-nilly.

In other news, I'll upload what I have so far in my spreadsheet later today with a large spread of WD values at 2 different AP points, both at 298 Determination. It also has that stupid potency thing I showed earlier where everything is stupid. >_>

[SunnyHirose]

Yeah, I'm not saying to ignore outliers. Answers are often found in outliers. That one min/max point is the most extreme aberration I can recall, though. I can't easily explain that even on a theory I've been trying to test, one where 150 might get rounded down (on the basis that it could get rounded to 142 and 157 whereas the other two potencies go evenly into 20).

[Viridiana]

Yeah, I'm not saying to ignore outliers. Answers are often found in outliers. That one min/max point is the most extreme aberration I can recall, though. I can't easily explain that even on a theory I've been trying to test, one where 150 might get rounded down (on the basis that it could get rounded to 142 and 157 whereas the other two potencies go evenly into 20).

And then you have the case where the 150 potency is disproportional in the opposite direction from above.

           150 Min 150 Max 200 Min 200 Max 320 Min 320 Max
72WD 803AP   504     556     670     741     1072     1185

Also, spreadsheet as promised.

Edit: Yellow blocks in the spreadsheet are values that are probably accurate min-max, but may be off by 1 in one of the directions.

[SunnyHirose]

Your data are making me pull out hair; I'm used to recording every hit for some more in-depth statistics :I

I can come up with a zany but plausible double-rounding scheme where a skill's potency is translated into an integer damage and then given the ±5%, but that doesn't prove to be anything like a solution for these strange results, because even if that's what happens, the numbers are still too far off for it to work.

[Viridiana]

Your data are making me pull out hair; I'm used to recording every hit for some more in-depth statistics :I

I tried that once, but having to individually label each set for WD/AP/Det got so tiresome so quickly that I just gave up on it.

[Kenji1134]

SunnyHirose, this data may be useful to you.

https://drive.google.com/file/d/0B7E...ew?usp=sharing

It's fairly well organized so you shouldn't get too lost.

I will say as a forward, I personally believe that 1 pot = 1 pot, end of story. It makes no sense from a programming standpoint to make 1 equal to 1.00X, where X can change slightly for some classes, and falls WELL within the statistical variation imposed by the +- 5% swing... It also makes no sense for a model to have any negative coefficients aside from an intercept if there is no (Stat - Base_Stat) variable in the model... Having a stat lower potency in any way is dumb... Moving on...

The BLM data and the 4 sets of WAR data are from just wailing on the dummy, collecting the results in ACT, removing crits, and dividing the average of the set, which at about 60-70 hits has a standard deviation divided by average damage of roughly 2.5%... so its statistically good enough.
And then divided that average damage by the ability potency and all applied buffs and traits.

But what you may find MORE useful is the 2nd set of WAR data on the 1st sheet.
I managed to set up my gear and stat allocation such that I could change 1 stat (WD,Str,Det) at a time. So this is a whole lot of single variable adjustments.
Additionally, Flash, which has a base potency value of 600, does not have the 5% variance that everything else does, it produces the same amount of hate EVERY. SINGLE. TIME.
Combine that with Maim and Defiance (which now has a hate mod of 2.3x up from 2.0x), and we can test with a 1242 potency ability which only produces 1 value for a given set of stats.

The other sheets are my own attempts at determining the ideal model, or rather the "true potency formula".
The problem with these regression models is that they are only valid for the data fed to them. So if someone who was i200 or i210 were to plug in their stats and potency value, the formula would probably be off by a few %.

I suppose the ultimate goal will be to find the form of the damage formula, rather than the coefficients. If the form is correct, I mean 100% correct, with some equation massaging Excel can spit out the coefficients pretty easily if someone's willing to parse for a while.

My theory is that there are 3 "real" equations.
Caster Spells and Heals
Abilities for Classes with Autoattack
Autoattack

Also to Dervy, tell me how this fits your DRG data: Potency_Value = 1.74765e-6 * STR*WD*DET/218 + 4.37529e-5 * STR*WD + .001023549 * STR. It fits Warrior amazingly well.
If this particular formula fits DRG, MNK, BRD, NIN, etc within less than 1%, ideally less than 0.5%, then we may have 1 of the 3 major potency equations for the game... at 60 that is.

PS. So this is my 2nd time doing this. The 1st time was with Purostrider over a year ago now. And as "fun" as the pursuit of a functional model from which we can derive the bountiful fruit of maximizing our dps and dragging raid content through the dirt is... I really wish a dev would pop in, say "You guys did a great job and we are honored by the amount of effort and love you have put into our game. And since you already got so close, here are the actual formulas that we use in the server-side code!"

That'd be nice wouldn't it? =)

[SunnyHirose]

MFW

Yeah, finding the form for real and not getting confounded by these potency and rounding shenanigans are the main reason we're even looking at said shenanigans right now. Unfortunately, if Viridiana's data hold true, we are well and truly in trouble with any assumption that potency has a proper direct linear relationship with min/max values in the final output. And that's even where it is possible to distinguish between weaponskills, abilities, etc., as some of these odd measurements are weaponskill-to-weaponskill comparisons and satisfy the full positive/negative spread.

We already know the game straight up isn't telling us about the real base value of stats like crit and skill speed, so how this bodes for Determination is a mystery to me. I could test while I level something but... man, it's such a chore

I don't know where those enmity values come from, but you may be interested to know that there's one other place I know of where the game actually gives an exact, non-varying value on player damage, and that's the damage of a non-crit DoT tick. I'll investigate that sometime, I think... if no one beats me to it (oh darn!).

Well, if I do have a suggestion for anyone looking to use a solver: you might want to try biasing the damage values up. The game almost certainly rounds down.

[Dervy]

Well, if I do have a suggestion for anyone looking to use a solver: you might want to try biasing the damage values up. The game almost certainly rounds down.

I already do that. It's improved the Accuracy of my Damage Model and Skillspeed DoT calculation by a heck load. For reference, my Variance of Errors when down from 3.6 to 2.5 for my WD Solver, and 4.2 to 2.13 for Skillspeed.

Also to Dervy, tell me how this fits your DRG data: Potency_Value = 1.74765e-6 * STR*WD*DET/218 + 4.37529e-5 * STR*WD + .001023549 * STR. It fits Warrior amazingly well.
If this particular formula fits DRG, MNK, BRD, NIN, etc within less than 1%, ideally less than 0.5%, then we may have 1 of the 3 major potency equations for the game... at 60 that is.

And uh, Unless I'm doing something wrong, it's overestimating my values for some reason, which is odd, as Dragoons should have a higher base Potency than a Warrior.

[Viridiana]

And that's even where it is possible to distinguish between weaponskills, abilities, etc., as some of these odd measurements are weaponskill-to-weaponskill comparisons and satisfy the full positive/negative spread.

To make a list for the data I gave earlier…

Dots: 30, 40
Ninjutsu: 180, 360
Abilities: 80, 100, 140, 200 (Assassinate), 500, 300 (incomplete and listed to the side)
Weaponskill: 60, 120, 150, 200 (Gust Slash), 220, 240, 260, 280, 320

Dots were done with 0 SS, of course. Also I should note that I don't recall ever seeing any variance between WS and Ability potencies, but it's not something I've ever searched for.

[Hustensaft]

I suppose the ultimate goal will be to find the form of the damage formula, rather than the coefficients.

If you want to find the formula - you just can't start at X, you have to start at zero, i.e. break your weapon, reset your 35 bonus points, write down AA, ability and heal potency for all classes and compare them. Then try to to find models and count how many you need to fit all of it. (In 2.0 one linear model did predict AA, heal, ability, and spell damage values for fully naked jobs over STR/DEX/MND)

Fun fact: Taking my level 50 warrior - fully naked auto-attack values are actually 20% higher in 3.0 compared to 2.0 (no - it's not maim)

From there you go by adding WD, Primary or DET.
Then combine either of those.
Then combine all of them.

We already know the game straight up isn't telling us about the real base value of stats like crit and skill speed, so how this bodes for Determination is a mystery to me

True, true!

Since obviously CRT and SS do nothing at their respective base value, the best thing to assume with practically no way to prove it is that STR and DET do stuff linearly from 0 to StatValue.