Greedy Goblin

Monday, May 27, 2013

I told you it's not normal distribution!

Few things annoys me as much as claiming that the results of people show a Gaussian distribution: most people have average results and as you move to the extremes, you get less and less people. There is no doubt that the abilities of people follow a Gaussian distribution: our IQ, our strength follow it. In games the amount of time real life allows us to play also. The "results are Gaussian" claim is merely saying "your results depend on your resources". If you play better, you surely played more. If you have better job, you were lucky to have high IQ and a high-earner family to get you to college. Saying "results are Gaussian too" is saying "your results are out of your control".

To disprove it, I started to collect results. It wasn't easy to get unbiased, straightforward dataset. The idea came from my girlfriend: Megaera. This is a very simple raidboss in World of Warcraft for damage dealers: stand where everyone else, shoot what everyone else, minimal "dancing" and even that is ignored in LFR. The Patchwerk of Pandaria. So I resubscribed WoW (just for this project, don't dream EVE players!) and went to kill Megs. 25 times in the same lockout (the week before the patch):
My girlfriend donated another 6 dead Megs so I had 31 screenshots of damage meters. Since I found that in most raids one damage dealer was AFK/DC/offline, I ignored the 17th, so got 16 data from each kills, 496 data points all together. Since the "DPS" field of the damage meter is weird, often providing higher number for a player who did less damage, I simply divided damage done with fight length (and 1.05 if there was a wipe before).
This is probably not an accurate DPS as there is time spent running between heads, but I don't care about that, I wanted to compare performances. So for accuracy: the below data is exact for "damage done on Megaera on LFR" and approximated for DPS.

With that said, the average, rounded DPS of the dataset is 65K with the sigma of 20K. I divided the players into half-sigma (10K) wide groups, so the middle group with the name "65" means those who had between 60 and 70K DPS. The left column is to be read "between 0 and 30" instead of "between 20 and 30" while the rightmost is "100 and above". The results are these:
As you can see, the Gaussian doesn't fit the results well, while "DoubleGauss" do. For easier comparison, here are the error bars (difference between real and fits). As you can see the "DoubleGauss" has much smaller errors.
The mean square difference is 0.0183 for Gaussian and 0.0078 for this "DoubleGauss", with most of the error coming in the two rightmost bars. I think it is because of the set bonuses, if someone gets good gear, he gets disproportionally better stats via set bonuses, so there are less "almost very good" players than "very good" ones. Both fits fail similarly for these bars, if we ignore them, the mean square differences change to 0.0168 and 0.0035 respectively, making this "DoubleGauss" almost perfect fit, within the granularity range (you can't get smoother results than 1/496 if you have 496 exact datapoints).

What is this "DoubleGauss"?
It's the sum of two Gaussians, the blue has the mean of 69K with 20K sigma, the brown has 47K mean with 10K sigma. These are two different, overlapping Gaussians, products of two yes-or-no different groups. It's like trying to measure size of dogs and figuring out that some of our subjects are actually cats. While there are genuine small-sized dogs (blue is non-zero in the sub-40K DPS), most of the low size results came from the invading cats.

This is the point when - instead of being satisfied that I was right - I double-checked everything because I did not want to believe it: the brown "scrub-player" Gaussian has 21% weight only. I - along with most good players - expected to see a small "elite" next to the large group of "n00bs". Instead I found that the good and average players belong to the same dataset, and the scrubs are the small minority that doesn't belong.

This is probably while Blizzard is losing customers: everyone - including myself - believed that the majority of the players is bad and catering to them is a good idea to make the game popular. But no, they are a small, but vocal minority. When Blizzard dumbed everything down to their level, they served this minority against the majority. While there are bad players who belong to the majority, they are indeed newbies and very low playtime players, who are like us, therefore don't need freebies, just shorter timespan content and more materials to learn.

The above graphs don't tell who are the "scrubs" besides two facts: they are unlike the majority and they suck at WoW. Tomorrow we'll see who they are.

8 comments:

Rohan said...

Very interesting results. Could you provide your raw data, maybe as a Google spreadsheet?

Gevlon said...

If I had any idea how to do that, I would.

However I can e-mail it to you, it's not that big.

Aziraal said...

Quite good analysis.
I've myself reloaded my wow account few month ago and i feel the same but it's because of the human nature itself.
When you considere yourself "at the top" you only notice (or maybe you notice far more) the noob people and the weirdest situation.
What makes you think there are a lot of noob is not an average distribution, it's a particular situation when you get 50% of your lfr filled with noob.
You point the noobiness of people in lfr by it's far far worse in PvP.
Lot of people imagine they are good simply because they make lot of kill/damage (even the good stuff one) and though lose the bg quite fast cause 50% of the team isn't playing for the objectives.

Alkarasu said...

@Gevlon
"If I had any idea how to do that, I would."

1) Go to https://drive.google.com
2) Upload your spreadsheet file (it understands most common formats)
3) Open it from there in Google Docs
4) In the upper right corner there will be big blue button with access control
5) Change the access level to "Anyone" or "Anyone with the link" (it's "Personal" by default). The link itself will be on top of the window.
6) You are done, your spreadsheet is avaliable read-only to anyone you provide with the link. Works with most office document formats as well.

Alrenous said...

DPS: last I checked, the meter personalizes fight lengths. I think it assumes that if you're asking about DPS, you want to know who will usually do more damage, not who did the most this particular time. The latter is already listed.

-

I'm weakly confident there's a third, tiny Gaussian in the 100+ range, part of a group that almost never goes to LFR.

The reason it appears to you that almost everyone is a moron or slacker is that you're in this third group. From this vantage, the middle group looks pretty bad. As bad as the bottom group looks to the middle. It's easy to tell them apart in practice, because they're culturally different. Attitudes, usage of 'lol' and so on.

Or so the prediction goes. I'll see if it stands up to the next post.

Camo said...

Quite interesting, but you have to keep in mind for whom LFR is made: those who can't play in normal or even heroic raids.
I'm not saying they will never go LFR but those players are present until they get the items they need and never come back.
You can observe the same on reset day when the better players queue and you progress at a higher speed and less wipes than the rest of the week.
Another problem is of course LFR itself. It doesn't require your best so you can slack and leech, and many do.

Anonymous said...

Did you attempt chi-square test to prove your hypnotises? Dataset seems big enough for that.

Rohan said...

Sure, email me the raw data if you can. My address is rverghes@gmail.com.