How to analyze loot

[Doer]

Originally Posted by Dawis

1. Although there is strong corelation between monsters health and size of it`s loot we loose much of this corelation if we look only at globals we get from this monster. Why?

Because these are nearly exponential distributions and there is a lot of data, he can still fit the "tail" of smaller mobs (e.g. argo young) as well as for bigger mobs. I don't think that particular issue is a problem. Then again, maybe i'm missing something.

[Dawis]

Originally Posted by Doer

Because these are nearly exponential distributions and there is a lot of data, he can still fit the "tail" of smaller mobs (e.g. argo young) as well as for bigger mobs. I don't think that particular issue is a problem. Then again, maybe i'm missing something.

Well - take simple average of Argo globals and simple average of Aurli globals.
Aurli global size will of course will be higher, but not much higher. Although Aurli has many times the health of argo the average global will be reduced by lots of small globals.
Let me elaborte:
Lets assume loot function
Loot = Health * C(x)

Lets assume C(c) generates constants from unknown parameter vector x from 0 (no loot)to very high constants (ATH)

Argo case Health = 200 --> loot = 200*C(x)
Aurli case Health = 2000 --> loot = 2000*C(x)

Remember that actualy we want to observe properties of c(x) function not whole loot function.
Of course in second case loot will exceed 50 ped much more frequently.
To compare loot is size with respect to monster you must divide it by HP (which strongly corelates to cost to kill that mob)
so 50 PED argo = 50 / 200 = 0.4 = C(x)
50 PED Aurli = 50 / 2000 = 0.04 = C(x)
So in aurli case we would take average from C(x) results in range 0.04; inf
but in Argo case we would take average from C(x) results in ragne 0.4; inf
That is big difference.

[falkao]

Originally Posted by Dawis

Hi all.
..
1. This is not an attempt to offer real loot function underneath the EU, this is simply statistical research of results this function gives - thus real loot we get over time. Which is ok - you can`t prove any reasonable theory about loot function anyways - so lets just analyze what is coming out of the black box.

that's correct and I stated this in the text. Nevertheless, there will still be one confusing it.

Originally Posted by Dawis

2. It is a bit difficult to get large samples of data even for results of this black box - loot we get. Each avatar can only get full data about his own results - and even that is not easy - so more acurate data we want - more difficult it is to get in large enough samples.
We can escape this problem, by reading only globals that appear into the system - because system offers a nice way to get access to data about all loots above 50 ped. (big respect to person for building the http://87.62.217.174/login.asp system which allows everyone to access this data easy)
As I understand author of this thread uses results of this system for building his theory.

indeed, and for an exp dist you need quite some data to get all aspects. It would be easier with a normal dist.

Originally Posted by Dawis

...
One easy way to do this is assume Loot = health * C(x) where C is constant generated at event x by our black box of loot function which inner workings is not our concern.
This provides us with easy solution of only measuring Loot / health not plain loot.(we assume loot proportionaly increases by increasing monsters health)
Thus by measuring Loot/health we measure how good is loot with respect to monster we are hunting.
As I said I did not have time to understand all maths behind authors theory, but I believe author has taken health into account correctly.

thats a good approach and one can analyze loot that is health corrected. However, you loose the possibility to model health, since you corrected for it. I did this analysis in the beginning to understand what kind of dist we have.

Originally Posted by Dawis

1. Although there is strong corelation between monsters health and size of it`s loot we loose much of this corelation if we look only at globals we get from this monster. Why?
Simply: If monster monster A has 200 health and monster B has 2000 health it does not mean that every global you get from monster B will be 10 times as big. Not even close, becasue you will get a lot small globals from B, as oposed to only rare globals from A.
This raises a problem, because current method uses data we get from globals. Only possible solution I currently see is - increase minimum global date we take into account from monster B proportionaly.
Let me explain:
1. Killing 1000 monster A you will probably get 2 globals
2. Killing 1000 monster B you will probably get 20 globals

Of course you will get lots of small globals from B - because loot proportionaly increases with size of monster hp. So to save corelation between monsters health and monster loot in B case we look only at the globals which exceed 500 PED.
Math nehind this is easy:
Monster A health 200
Monster B health 2000
Ratio = 10
Assumption - loot should be 10x greater from B

We cant get acces to all loot date - we get only global data. So minimum loot we see is 50 ped.
In A case this means we measure only the best loots from A monster (pikes of loot function).
In B case this means we measure not only the best but medicore loots to.

To compensate this we multiply minimum measurement requirement by 10 in B case. So we get aproximately even amount of data samples in both cases.

Ok.. I am exausted for now

fortunately a truncated exp will have the same characteritics as one that is not truncated. The only difference is the different mean you'll get. If you know the truncation point the original mean is estimated mean - truncation point. However, it is also possible that globals start at 50.

In analyzing globals you won't get any information on normal loot. We had some threads about that before and both results should be finally combined.

I used the Cos model to test for health on loot. The advantage with this model is, that you don't have to know the real distribution.

You'r porposal to analyze only loot higher than some threshold is dangerous. As outlined before, we do observe a mixture of several distributions. Therefore in using a threshold you'll hardly know what kind of data you have. You have to know one of the distributions to find a good threshold.

[falkao]

Originally Posted by Dawis

Well - take simple average of Argo globals and simple average of Aurli globals.
Aurli global size will of course will be higher, but not much higher. Although Aurli has many times the health of argo the average global will be reduced by lots of small globals.
Let me elaborte:
Lets assume loot function
Loot = Health * C(x)

Lets assume C(c) generates constants from unknown parameter vector x from 0 (no loot)to very high constants (ATH)

Argo case Health = 200 --> loot = 200*C(x)
Aurli case Health = 2000 --> loot = 2000*C(x)

Remember that actualy we want to observe properties of c(x) function not whole loot function.
Of course in second case loot will exceed 50 ped much more frequently.
To compare loot is size with respect to monster you must divide it by HP (which strongly corelates to cost to kill that mob)
so 50 PED argo = 50 / 200 = 0.4 = C(x)
50 PED Aurli = 50 / 2000 = 0.04 = C(x)
So in aurli case we would take average from C(x) results in range 0.04; inf
but in Argo case we would take average from C(x) results in ragne 0.4; inf
That is big difference.

here is box plot comparing argo y vs aurli devast


Click to enlarge

the difference is highly significant. Not sure if this answers your question.

[Dawis]

May be I can use analogy to explain.
Suppose you wanted to listen to 2 radio stations and measure how strong signal per average you can receive from them.
Station Argo is located very far, while station Aurli is located very near.

Therefore signal strength from Argo station is reduced ten times when it reaches you, wher as signal strength from Aurli stays the same.

Of course there is a minimal signal strength which your antenna can recieve, and register.
Suppose you register all signals sent from Aurli and Argo.
What you will get - From station Argo you will receive only those signals which were originally the strongest - those signals that exceed your antenna minimal signal by at least 10 times. From station Aurli you will receive almost all signals that exceeds your antenna minimal signal strength.

Now if you simply take the average of signals received you will not get a nice picture of what is signal strength coming from Aurli and Argo.

Signal coming from Argo will of course appear weaker that signal coming from Aurli. But much to the surprise signal coming from Argo will not be simply 10 times weaker - because your antenna has minimal signal strength it can recognise - you will lose a lot of small signals comming from argo.

So you can`t use your antena to compare average signal strength.
However you can measure and compare both stations strong signal average.
Those signals that are above some margin - In this case 10 times as strong as your antenna minimal signal strength recognition - because you wont register signals bellow this strenght coming from Argo station.

[Dawis]

The picture is nice. But I am not sure I understand how to read it correcly.
I did notice, however, a biger box for Aurli devastator.

But my basic idea: Argo 200 HP, aurli devastator 3k HP.
So it costs about 15 times as much to kill one Aurli devastator than it costs to kill one Argo young. So normally one would expect 15 times as much loot from Aurli.

If loot is more than 15 times greater from Aurli then one can say Aurli devastator loots better than Argo young.
If loot is less than 15 times greater from Aurli then one can say Aurli devastator loots worse than Argo young.

Now tell me what your theory says:
how many times greater loot will you get from Aurli devastator than from Argo young per average? (I am not sure I can read it from chart)

[falkao]

some experimental stuff added to main post.

[falkao]

Originally Posted by Dawis

The picture is nice. But I am not sure I understand how to read it correcly.
I did notice, however, a biger box for Aurli devastator.

But my basic idea: Argo 200 HP, aurli devastator 3k HP.
So it costs about 15 times as much to kill one Aurli devastator than it costs to kill one Argo young. So normally one would expect 15 times as much loot from Aurli.

...

your observation is correct and your example with radio stations works quite well. However, statistically this means that to estimate parameters for an exp dist with higher mean you'll need more data to get the same accuracy as when the mean is lower. So for Argo I need less data as with Aurli.

The depicted boxplot shows the median in the middle and the interquartile range. Since the standard deviation for an exp dist is the same as the mean, Aurlis will have a larger box.

Originally Posted by Dawis

..
If loot is more than 15 times greater from Aurli then one can say Aurli devastator loots better than Argo young.
If loot is less than 15 times greater from Aurli then one can say Aurli devastator loots worse than Argo young.

Your proposal seems reasonable but is unfortunately wrong. Proportionality doesn't mean that.

Please read the newly added "experimental section". There I describe that mean loot might depend linearly on hp. So we have

m Argo = a*300HP + b
m Aurli = a*2800HP + b

For the first mean this leads to a factor of about 4. (a*x1+b)/(a*x2+b) can't be further simplified. Btw, mean for Argo in my dataset is 90 and that for Aurli is 166. The proportion leads to a factor of 1.8. That quite different from 4. The reason for this is, that we have 3 means to consider per mob and that my sample size is to small to cover all aspects. However, I'm at the beginning to analyze this, so I can't conclude more atm.

[Legion]

Falkao check you pm's :P

Also i dont think the globals are exactly erm exponential(?)

Because if you hunt bigger mobs you have a bigger chance to profit but at the same time a bigger chance for loss sort of.

[Dawis]

Originally Posted by dawis

Originally Posted by Dawis View Post
..
If loot is more than 15 times greater from Aurli then one can say Aurli devastator loots better than Argo young.
If loot is less than 15 times greater from Aurli then one can say Aurli devastator loots worse than Argo young.

Originally Posted by falkao

Your proposal seems reasonable but is unfortunately wrong. Proportionality doesn't mean that.

Well I am sorry - I totally missed that Argo young has 300 not 200 HP.
But other than that I don`t see that I am wrong.
If killing one Argo y will cost you X PED, then killing 1 Aurli d will cost you 10x PED. So normal loot from Aurli would be at least 10x as much as Argo.
Of course we can`t use any of these monsters as basis for "normal loot".
But If loot from Aurli is more than 10x greater we can say that Aurli loots better than Argo. Otherwise we would claim that Argo loots better.
where am I wrong about this?