Get an Epic Experience with Premium
• posted a message on Magic Find and its efficiency: A statistical insight
Yeah, I've done exactly what you described, I can upload my calculation if you still want to see them, but it's getting quite late here so this can wait till tomorrow. Due to my background I'm used to making 1sigma error bars, you seem to prefer 95% confidence levels? That's the simplest explanation I can some up with :P. Otherwise are you using the combined data from all the acts for the item abundances and from all the MFs for the ilvl distribution?

The idea of item cut offs is actually quite interesting. I don't think it can reproduce the step at 55, but I'll definitively keep an eye out for ilvl52 polearms next time I'm farming.

The only reasonable model I can think of that would naturally produce the 55 point would be if we had e.g. 4 treasure classes 51-54 armor, 55-59 armor, 51-55 weapons, 56-59 weapons. This would produce the step as observed with an intermediate point at 55. Such an item classification seem pretty arbitrary though.
Posted in: Theorycrafting and Analysis
• posted a message on Magic Find and its efficiency: A statistical insight
Quote from ztking

Interesting... did you include normal (white) items only or also legendary?
I ignored legendaries for this analysis.

@Nubtro: If you allow all affixes to rolls with a different probability, the situation becomes very complicated. The problem is that if you want to calculate the probability of an item having a certain affix A, the first roll would have the probability pA. Now if the item rolls a different affix B, the chances of rolling A on the next roll are increased (since B is not an option anymore) by an amount dependent on pB.
This basically means that the probability of observing the affix A on your gloves depends on the nominal chances of all other affixes as well and I actually doubt this has an analytic solution at all. If you want to get accurate results your best bet is probably to make some simulations. Otherwise n/ (150*6) is probably a pretty good first-order approximation.

I don't know if yo're aware of it, but there's a list of all possible affixes here: http://d3inferno.com...temAffixes.html
I haven't looked into the list in too much detail, but could it be possible that there are simply more different affixes that give you allres then indestructibility so that the nominal chances of all the affixes are actually identical?
Posted in: Theorycrafting and Analysis
• posted a message on Magic Find and its efficiency: A statistical insight
The following analysis will assume that the item type is determined before ilvl.
The observed ilvl distribution is then:
p(L) = sum_(item types with ilvl=L){ p(type)}*W(L)
Here p(type) is the probability of a certain item type dropping and W(L) is some ilvl distribution that is used for the ilvl roll. In order to calculate W(L) we can divide the observed ilvl distribution by the sum. For the item level distribution the observed values are used without any further assumptions. The resulting W(L) distribution can seen in the attached plot.

The error bars represent uncertainties due to limited sample sizes.
The interesting thing in the distribution is that it's almost constant between 51-54 and 56-59, there's a jump at 60 and it seems to decrease linearly afterwards. I think this strongly suggests the existence of certain ilvl groups (Treasure classes?). The only problem is the point at 55 which seems to lay between the two groups.
Posted in: Theorycrafting and Analysis
• posted a message on Magic Find and its efficiency: A statistical insight
@ztking: I'm sorry. I'm following your logic, but I'm getting to the opposite conclusion than you. If ilvl rolls last (as you concluded) than you would see no difference in the item drop rates between different acts, which we apparently do observe.

Edit: haven't seen your edit, it does make sense now

Edit2: I think the biggest problem with the proposed system id that it poorly explains the observed ilvl distribution. Yeah it's possible that the distribution is coded like that, but I think it's very unlikely. Or are we still talking about ilvl groups? So it would be more like: Quality --> ilvl_group --> Item Slot --> ilvl
Posted in: Theorycrafting and Analysis
• posted a message on Magic Find and its efficiency: A statistical insight
Quote from ztking

EDIT: HEY! Anyone else notice that the chance of jewelry, follower special, and class specific armor (which all has a max ilvl below 63) takes a dip in Act 3? I think this is strong evidence that ilvl is determined before item type, else we'd be seeing the same number of those items but their ilvl distribution would be different.

Yeah, this does look a bit fishy. The statistical significance of the difference is still a bit low though. Also coincidentally most of non63 items are also magic+ items. I can't think of a mechanism that would make magic+ items act dependent, but it's definitively something worth keeping at the back of the mind.

P.S.
I extended the magic+ rows in the spreadsheet to get values for different acts since this might provide some interesting insight as well.
Posted in: Theorycrafting and Analysis
• posted a message on Magic Find and its efficiency: A statistical insight
Just for clarification, I'm getting confused by the nomenclature at times; what is a Treasure class? set/legendary/6/5/4affix rare...?

Quote from ztking

Edit 4: I also HIGHLY doubt that Treasure Class is determined after both Item Type and ilvl. If a lvl52 necklace was chosen, it would be forced to be a legendary, so you'd have to play with the ilvl roll to make that happen less often, which would require a special distribution of ilvls for each item type that accommodates whether it's always "normal", always legendary, or could be either.
This still leaves the possibility of:
Item Type --> iLvl_grouped --> Treasure Class --> Choose Available Item
As long as you choose appropriate ilvl groups you can get rid of the problem of forcing an item to be legendary.

Another think we could exclude is quality(treasure class) rolling before item type. Since MF has an effect on the quality roll and not all items are present in all qualities, MF would have a noticeable effect on the item type distribution. I did a short test on this a few pages back and did not find such a correlation. (Although some more data might be needed for definitely excluding it.)

I think so far the most probable options are:

Item Type --> Treasure Class --> iLvl_(flat/grouped) --> Choose Available Item
Item Type --> iLvl_grouped --> Treasure Class --> Choose Available Item
Item Type --> iLvl_grouped --> Choose Available Item

I don't really like the last option since it would make MF act directly on the "choose available item" roll. Although this is possible it seems to make the process much more complicated than if we have a separate roll for it.

I think the main question at the moment is whether there are ilvl groups and if so, what are they.
Posted in: Theorycrafting and Analysis
• posted a message on Magic Find and its efficiency: A statistical insight
Quote from Lexikon

There is an OCR project at inDiablo. The released version just cuts screenshots, but the most recent version in development can identify items, I think.
I'm so glad to hear that. I was trying to make it work today, but it's still missing way too many items. I'll take a look at their project, I'm sure we can use it somehow.

Regarding the drop data (3462 drops so far): Even if item type rolls first they're not necessarily round numbers. I'd try to figure out some ratios between the drop rates. E.g.
• All armors (+shields) have the same drop chance. We can use this as the base drop rate.
• All class specific armors have the same drop chance. The total chance of a class armor dropping is equal to the base chance.
• There's a base chance to find a ring/amulet/follower item. The chance for rings is twice as high as for the other two types.
• Weapons have half the base drop chance.
For class specific weapons things look extremely weird. I'm really baffled by the large variation of the drop rates of weapons for different classes. I mean it's not like monks would need more fist weapons than barbs mighty weapons???

EDIT: Are we sure the played class has no effect on the drop rate of class specific items? Just for reference I got maybe some 200 items with my WD.
Posted in: Theorycrafting and Analysis
• posted a message on Magic Find and its efficiency: A statistical insight
Don't know how well this is known, but mighty belts apparently can only be magic+ as well. I guess they should be considered jewelry for our purposes as well.
Posted in: Theorycrafting and Analysis
• posted a message on Magic Find and its efficiency: A statistical insight
@cyeron: It seems like a really neat idea to me, but honestly I find your chart quite confusing as it is.
E.g. item quality and item rarity should be the same thing? What's exactly the item integer/ how is it different then the number of affixes? Also I don't think we have enough information to make a definite conclusion about the slot/type/level relation.
Posted in: Theorycrafting and Analysis
• posted a message on Magic Find and its efficiency: A statistical insight
@ztking: what are the confidence intervals in your spreadsheet? I was assuming one sigma, but it doesn't really seem right. 90% confidence 95%?
Posted in: Theorycrafting and Analysis