Very nice. Have you considered testing out other bosses from Act 4? I'd be curious to see drop rates on bosses that have specific look, like Kaom's.





I tried to find the true "C" equation, given the IIR IIQ values you provided, and the subsequent actual multiplier to the amount of rares and uniques received from the control group (where you have no IIR and IIQ bonuses).



For ease of math, assume when i say IIR and IIQ I really mean (1+IIR/100) and (1+IIQ/100) respectively



I am taking a regression approach to this--- problem is there are so few samples! only 13! So I had to make a very simple model.



The model proposed in the original post has a large error to the actual C values.

So for reference, the average error of the original equation is:

Error = 4.218721564.



Let's see if I can do better.



I decided to take three different models and share the results.



FIRST APPROACH: explicit regression equation <Math Alert>

This is the "if I were programming it, how would I have done this" approach. And I've found some evidence I may be correct.



My inspiration came from this post



This lead me to believe the equation should look like . where f(IIR) is some function of the IIR.



I thought immediately that if I were to program it-- i would keep the ratio of rares:uniques the same, but start decreasing the amounts of whites/blues (depending on the monster maybe blues:rares:uniques ratio would also be maintained)



To test this thought I did linear regression on the ratio of rares:uniques. Where x=C values (your model) and y=averageUniques/averageRares



graph is here (big)



As you can see-- there is no positive trend between "C" and the ratio of uniques to rares! so now I can propose my equation.





Here, let p = % of drops that are rare or unique before any IIR or IIQ bonus. (IGNORE THE tick, my bad)





Basically what i'm doing here is lowering the amount of whites, and increasing the amount of rares. This limits IIR so that bigger values will have larger diminishing returns



Also notice: WE DON'T KNOW P!! and it might be different depending on what you're fighting!



So I just did the regression on p, since it's the only unknown



This is the "if I were programming it, how would I have done this" approach. And I've found some evidence I may be correct.My inspiration came from this post https://www.pathofexile.com/forum/view-thread/28697#p414204 . "Increased Item Quantity[...] gives more items.More drops means more chances to roll currency instead of gear."This lead me to believe the equation should look like. where f(IIR) is some function of the IIR.I thought immediately that if I were to program it-- i would keep the ratio of rares:uniques the same, but start decreasing the amounts of whites/blues (depending on the monster maybe blues:rares:uniques ratio would also be maintained)To test this thought I did linear regression on the ratio of rares:uniques. Where x=C values (your model) and y=averageUniques/averageRaresAs you can see-- there is no positive trend between "C" and the ratio of uniques to rares! so now I can propose my equation.= % of drops that are rare or unique before any IIR or IIQ bonus. (IGNORE THE tick, my bad)Basically what i'm doing here is lowering the amount of whites, and increasing the amount of rares. This limits IIR so that bigger values will have larger diminishing returnsAlso notice: WE DON'T KNOW P!! and it might be different depending on what you're fighting!So I just did the regression on p, since it's the only unknown



final equation = IIQ*f(IIR) =



I found the best error was:

Average Error = 0.2961646339

Where p = .42 (see spoiler for explaination of p, should be adjusted by what enemy is dropping)

Much better than 4.128





SECOND APPROACH: What if i'm wrong about how IIR works?

What if it takes in two arguments, like f(IIR,IIQ)?

Well, then I should discover f(IIR,IIQ) through symbollic regression!



Symbolic regression is an evolutionary computation algorithm, basically designed to guess the equation to a set of numbers.



I used ECJ library to come up with the results, toyed with settings until i found something nice with low error.







fun fact: when I punished it for making strange equations like this (i had a measure for that-- too much detail for this post) -- i ended up getting my equation i proposed in the first approach :)



What if it takes in two arguments, like f(IIR,IIQ)?Well, then I should discover f(IIR,IIQ) through symbollic regression!Symbolic regression is an evolutionary computation algorithm, basically designed to guess the equation to a set of numbers.I used ECJ library to come up with the results, toyed with settings until i found something nice with low error.fun fact: when I punished it for making strange equations like this (i had a measure for that-- too much detail for this post) -- i ended up getting my equation i proposed in the first approach :)

final equation = IIQ*f(IIR) =



Average Error: .256 (WARNING: lower error not necessarily good! that equation is pretty strange, use as approximation only)



THIRD APPROACH: What if i'm wrong about EVERYTHING! (pure symbollic regression)



I just plugged it into ECJ and let it gooooo



Again, just because this has lower error doesn't necessarily make it better. This is because the more complex the model you have, the more it could "overfit" the data. Which just means that it may not generalize to new values! which is what we're most interested in here. We only have 13 samples, so overfitting is easy to do.





final equation =



Error = 0.114959915







Feel free to use which ever equation you think is best to figure out how you should assign IIR and IIQ (i'm personally a fan of the first equation)

I did some followup on your work. (if somebody has already done this... oops)I tried to find the true "C" equation, given the IIR IIQ values you provided, and the subsequent actual multiplier to the amount of rares and uniques received from the control group (where you have no IIR and IIQ bonuses).For ease of math, assume when i say IIR and IIQ I really mean (1+IIR/100) and (1+IIQ/100) respectivelyI am taking a regression approach to this--- problem is there are so few samples! only 13! So I had to make a very simple model.The model proposed in the original post has a large error to the actual C values.So for reference, the average error of the original equation is:Error = 4.218721564.Let's see if I can do better.I decided to take three different models and share the results.final equation = IIQ*f(IIR) =I found the best error was:Average Error = 0.2961646339Where p = .42 (see spoiler for explaination of p, should be adjusted by what enemy is dropping)Much better than 4.128final equation = IIQ*f(IIR) =Average Error: .256 (WARNING: lower error not necessarily good! that equation is pretty strange, use as approximation only)final equation =Error = 0.114959915Feel free to use which ever equation you think is best to figure out how you should assign IIR and IIQ (i'm personally a fan of the first equation) I double your bid.



IGN:paranoidcoder Last edited by paranoidcoder on Jan 31, 2017, 9:27:42 PM