General DiscussionSkill differences between players, which analytic function would best...

87 / 3 = 29 + 4 = 33.

C = 299,792,458 + 33 = 299,792,491.

The average MMR of DotaBuff is 5349.

299,792,491 / 5349 = 56046 and the average length of a game is 35 minutes.

1601 is the standard of deviation of dota 2's mmr curve.

Not sure what you are smoking, but I want some.

http://en.wikipedia.org/wiki/Normal_distribution

MARLAN

And with normal english sentences, how do you explain what you said above?

Krazy Kat

So you think the Normal distribution is a better description then log or exp? Why?

To me the logarithmic function seems most likely. Perhaps with an [edit: removed and and replaced it with an] exponential end for the 1000 best players. Or something like that.

To me the skill differences among the best 3-5 % of the players doesnt seem very significant. Therefore the logarithmic function seems likley since it increases very slowly in the end. And therefore it can be used to describ the small differences in skill flr the best players.

I am not talking about MMR. Then the ND curve seem good enough.

machine learning

You should start to explain what you mean by skill difference, I guess you mean their ability to dominate each other or their ablity to gain mmr ?

I doubt OP knows what he is talking about considering he had to look up the word "Hypothesis" and uses it wrongly.

Probably is a gaussian function, or a normal distribution.
Look at that plot (http://i.imgur.com/gWKPytL.png)

I just want to know why I have more than 60 % wins with KDA > 5 and still normal brackets.
I dont think im bad at all... rssrrs

Don't play your first game as support

Definition of "Skill" that was used in this topic: game knowledge and ability to adapt to the current game and situation

This is NOT = their ability to gain MMR. But it is correlated.
For example, I am better at getting the best stats in my team then increasing in MMR

The ability to win vs the other is not interesting to me.

#Road to 3k mmr

I did not look these words up. I know them all since my education. I copy paste them for the readers. Did you really not consider that?

Explain why I used the H-word wrongly?

I have studied Reasearch Design and Methods, but maybe you know this topic better then me?

It cant be the normal distribution since the skill should be on the Y-axis and the amount of players should be on a reversed X-axis

Atleast that is how I have visualised it.

I've always seen it as arcsine... massive amount of shit to learn at the start, massive difference between pubstompers and pros... just imo though...

this doesnt look as good as i thought

kappa

13oundary

I think many people have the same opinion as you. But I never seen any signs for why the end should be exponential (as the arcsin) instead of logarithmic. However, I did suggest a exponential curve for the top 1000 players. But that was more like an example and that is far away from arcsin since I suggest a logarithmic function befor that.

dango

Exactly. That is what i am talking about.

everything is a normal distribution if you are brave enough to apply central limit theorem

TripleSteal-

No. Not everything. You need high enough numbers. Otherwise the theorem is not valid. And why would 11.000.000 be enough players in this case? I dont know any reason that would suggest that.

And it also depend on the axises. Since I want skill on the Y-axis ND-distribution would suggest that the average skilled players are the most skilled players.

i know how its working lmao, dont explain me obvious things plx.
and actually i highly doubt that the normal distrbution can fit here. also, there is a huge difference in skill on very low (0-1500) and very high (5k+) levels, so i dont think log or exp functions are the ones you are looking for.

you can't quantify skill lol but mmr fits under normal distribution

u cant calculate cognitive skills, so there is no way to calculate skill difference between ppl.

There was already a topic and research done long ago by someone on dbuff on MMR statistical distribution using actual MMR values.

If i remember, it was a norm dist. Though admittedly, it should be more of a log-normal given that you cant go below 0 MMR and qualitative biases on disclosing MMR and sample framing.

Anyway, theres nothing more to discuss when op doesn't even have a metric for his definition of "skill".
Not that i know any better myself..

Ampoule Aquatique

Well, as with all scientific theories developed in the Universities, my goal is only to develop a model that is good enough. Not discovering the truth.

Mokujin

Ofc one can. Since the researchers can measuer happines, then ofcourse it is possible to measure skill in a computer game.

And I am not even try to MEASURE it. Just developing a theoretical model. You know, like a theory for the Big Bang without trying to measure the "bang".

TripleSteal-

I am expalining to anyone that reads this topic. It is not all about you, you know.

Well, since you dont think it is ND, LOG or EXP. Why dont you just say what you think it is instead.

Scrubby the Scrublord

"Anyway, theres nothing more to discuss when op doesn't even have a metric for his definition of "skill". "
This is just bullshit. I dont need a metric. I defined "skill" - that is enough. Enough for the purpose that I stated.

And once more, I am not talking about MMR.

I would have said if I had any idea, but I dont :) I only know how to criticize other ppl xD.

^a true dota player.

report OP pls kappa

There is no reason why you wouldn't be able to approximate player skill. We are doing far more complicated things in ml. You just need good representative data

omg nerds Kappa

I dont think that this guy understand that it was nerds that developed Dota1.

jussi

"We are doing far more complicated things in ml."

And what is ml?

You dont need data to develop a model. You need data to test the model.

here is my thesis. i think it's rly gut

^also note the "LEL"

Well it is nice that you show your art skills. But despite the intrinsic value in this creation, you are off topic.

^ok mate i will leave this topic thanks anyway

ML is machine learning.

"You dont need data to develop a model. You need data to test the model."

Here you are actually wrong. There are numerous ways of learning models/structures in data from just data. If you try to model player skill like people usually do then it will likely be biased in some way. Just look at that recent thing someone posted here where the algorithm awarded you extra points from just picking Troll.

jussi

What you are saying is that one can use data to develop a model. Which is true. But it is also true that you can develop a model without using data that was gathered in a systematic way. For example data in the form of knowledge and experience.

Yes, I know that. Those models have been proven to work pretty bad though

Well, then we have the same opinion. Which make you wrong when you said I was wrong :)

No. The models that you referred to have been working terribly. Ever couple of months or so someone comes out with a new MMR system that sucks balls. If even pro players can't reliably assess the skill level of a player why would some average dota player be able to do it? All these systems had major flaws.

decision making is one of the biggest factor in dota skill, so whatever your theory will be it will be at best 50% potentialy accurate if u knw what i maean. nyway i didnt sleep yet, nyway the point is im pretty sure its not worth losing ur time over such theory, cuz there iss no way for it to be accurate unless you have a machine that can measure intellectual skills and then bring those data into ur theory.

the point is there is just not enuff data to Really measure the skill

jussi

"Ever couple of months or so someone comes out with a new MMR system that sucks balls."

Please. for the fifth time or something. I am not talking about MMR.

"If even pro players can't reliably assess the skill level of a player why would some average dota player be able to do it?"

I am not average.

My closest people IRL think I am the smartest person they know. And they are not dumb. They have employments with salaies way above average.

"All these systems had major flaws."

So, what flaws do you see with the elements of the model I am trying to develop here?

Ampoule Aquatique

I am not trying to develop a theory. Only a model. I am not losing time since I am having fun right now.

F=m*a is not accurate. It is good enough during certain situations and for surtain bodies.
v=s/t is not accurate. It is good enough...

"the point is there is just not enuff data to Really measure the skill"

pff. How about you actually read my post before you tell me what I can do and not? I am not trying to measure. I am trying to develop a theoretical model.

Hmm this topic have been resting long enough.

Come on now guys and girls - show your intelligence!

The closest thing that comes to what you're describing is...

http://research.microsoft.com/en-us/projects/trueskill/details.aspx

Please stop littering the forums with your garbage, you're not a special snowflake you're as average as they come as Jussi said.

Take into account that both myself and Jussi have learned/are learning, and or working with modelling languages, machine learning, and similar stuff as are many others that frequent this forum. You're actually quite far off in what you're trying to achieve here.

You're basically just looking to come up with a model that assigns you as a "skilled" player though you have blaring lack of any type of skill.

"For example, I am better at getting the best stats in my team then increasing in MMR"

Like this game? http://www.dotabuff.com/matches/1432509657
Or this game? http://www.dotabuff.com/matches/1430988005
Or this game? http://www.dotabuff.com/matches/1430804469
Or this game? http://www.dotabuff.com/matches/1428643672

The list is never ending, garbage picks, garbage item choices, garbage playstyle, garbage theory crafting low skilled player who thinks he's above average and is a special snowflake.

Well I am on my phone so I am not going to copy paste the forum guide lines this time.

Who cares about ML?

I have an M.Sc. in engineering physics and one in innovation management.

From my educational perspective, this approch I am suggesting here is perfectly fine for the purpose that I atleast indirectly stated. It is standard research methology in these two subjects.

The LOG function is intuitivly reasonably and therefore a good starting point for the discussion. I am just interested in the theoretical discussion and the theoretical model. I have no intentions to go further then that.

You seem to assume that I actually care if people in this forum think I am a good player or not.

I dont care at all. I just care about the discussion, the intellectual challange and sometimes trolling.

I see no reason at all to care if people in this forum think I play good or not.

Remember, I am over 30 years and been insulted over 60.000 times during my more then 6000 games since 2007. If 100 people more in this forum tell me I am bad then it is just a small percentage increase.