A DPS Simulator 2 - SimFF

[Viridiana]

ACT does indeed.

Wait, it does? I must be blind. @_@

[SunnyHirose]

Wait, it does? I must be blind. @_@

I don't understand the format myself, assuming it's even visible. But the patch notes say so.

[Kenji1134]

Oh god, full ACT logs from a week's worth of raiding from every raider on every server...

I can feel that neural network having a stroke already. =P

But yeah, I understand that there is a Mathematica package which is good for modeling arbitrary equations from data. Sadly I am a MATLAB guy, not a Mathematica guy,

[pandabearcat]

It isn't too hard to do enough samples for testing a neural network.

Assuming we just want to do a simple static test, all we need are...

1. Any random combination of gear that gives a set of stats.
2. Do one known potency attack (lets do w/o damage modifiers for now).
3. Do that enough times to get an average output - alternatively, feed all into the network if we want randomness to be a part of it.
4. Repeat 1-3 with other potencies and random sets of gear.
5. As an addendum, for now separate crits and noncrits and focus on getting a good formula for noncrits.

Ideally all this data is collected on a level 50 training dummy with no other modifiers, because I am almost certain enemy level/defense modifies damage.

The neural network can be exceedingly simple - input layer with X inputs (corresponding to all stats + potency + level if desired), one hidden layer with some number, lets say X*2 nodes, and one output node being the experimental observed potency.

Really though not sure why this wasn't done years ago =P Much easier to do than trying to find a regression line in a spreadsheet - and guaranteed to be accurate.

EDIT: update on the sim, I'm on vacation until end of next week. Too much irl stuff (performance reviews lol) and need to take my cousin to college. If you haven't noticed already I've sort of lost my motivation for now, so hopefully I will have it back after a break.

[Viridiana]

Ideally all this data is collected on a level 50 training dummy with no other modifiers, because I am almost certain enemy level/defense modifies damage.

The level 1 critters outside Gridania still calculate the same as the dummies outside Idyllshire, though, so I kinda doubt that.

[pandabearcat]

The level 1 critters outside Gridania still calculate the same as the dummies outside Idyllshire, though, so I kinda doubt that.

Shrug. Fewer variables the better. Either way yall have all this data but I think are going the wrong way about getting info from it. Work smarter not harder!

[Kenji1134]

Assuming we just want to do a simple static test, all we need are...

1. Any random combination of gear that gives a set of stats.
2. Do one known potency attack (lets do w/o damage modifiers for now).
3. Do that enough times to get an average output - alternatively, feed all into the network if we want randomness to be a part of it.
4. Repeat 1-3 with other potencies and random sets of gear.
5. As an addendum, for now separate crits and noncrits and focus on getting a good formula for noncrits.

Ideally all this data is collected on a level 50 training dummy with no other modifiers, because I am almost certain enemy level/defense modifies damage.

The neural network can be exceedingly simple - input layer with X inputs (corresponding to all stats + potency + level if desired), one hidden layer with some number, lets say X*2 nodes, and one output node being the experimental observed potency.

Really though not sure why this wasn't done years ago =P Much easier to do than trying to find a regression line in a spreadsheet - and guaranteed to be accurate.

I linked an excel sheet on my google drive a few days ago in this thread, which uses the exact data collection methodology you described here.
So there's your data right there... Well, the War Flash data technically has Defiance and Maim in it, but they are factored out for the potency calc.

The specific attack used for the BLM data is mostly F1 under AF3, so 180*1.8 potency.
For the WAR damage tests it was an equal distribution of Heavy Swing (150) - Skull Sunder (200) - BB (280) potency.

So now all we need is said neural network. Happen to have one laying around? =)

PS. As for randomness, wasnt it confirmed (and likely still true) in 2.X that damaging abilities have a +-5% swing to them, and healing abilities have a +-3% swing?
To my knowledge the only things that are static are Flash and dots, though dots can crit. Flash cant crit.

[Dervy]

PS. As for randomness, wasnt it confirmed (and likely still true) in 2.X that damaging abilities have a +-5% swing to them, and healing abilities have a +-3% swing?

Yeh that still is the case. a WHM I'm talking to who's doing data collection for me atm confirmed it once more.

[pandabearcat]

I don't have time to tailor one...but they are so generic you can sit down with a book for 30 minutes and write one in the language of your choice.

Might (highly likely) have generic packages online as well.

If you like web coding (euuugh) you can prob just set up a gitpage with simple upload and dynamically update your model.

Again the only hard part is collecting the data and feeding it in.

The actual network is only a few (maybe a hundred) lines of code.

EDIT: good news! I've implemented macros. Just not...the macro grammar yet. Oops. Oh well, I'm feeling tired, time for a nap.

[Sidra]

Anything I can do on my SMN to help you guys? I can't program anything - thinking of in game dummy testing in whatever situations where I give you the data. I got ACT, and plenty of time.