20 replies to this topic

[Call_Me_Ishmael]

Greetings Pilots!

 

Some of you may know I've got a few degrees to my name, and have a inquisitive, mathematical bent.  This is what I've dug out of various online resources.  The Wikipedia implementation I did not pursue, but I believe is similar (but uses 350 as the maximum delta, which is not appropriate for a multiplayer game where a player often plays six rounds or more a day).

 

I would like you to think of this as peer-review opportunity - don't try to slam me as being clearly stupid, if you see something to improve, let's discuss it.  I can then modify the slides.

 

I can make the slides available (as slides) on Google Drive after a period of time.

 

My intent is to make it REALLY clear what MMR is and what it isn't, and let decisions be made with an understanding of the algorithm.  What I have here will be kinda-close to Hawken's actual implementation in effect.  I may have some details (tuning parameters) incorrect, but believe this is pretty damned close.  Elsewhere - NOT THIS THREAD PLEASE - is the place to discuss if it's too harsh. I'd like to use this thread as the peer-review of the algorithm.  When, as a group, we're happy with the understanding, I'd just as soon lock this thread (and publish the slides).

 

Best,

 

Lunatic_Fringe (aka Ishmael)

 

 

 

 

 

 

 

 

edited title to make it more appealing.

Edited by Call_Me_Ishmael, 27 May 2015 - 02:53 PM.

[plons]

why am i laughing ???

 

the matchmaking mmr and all that stuff was a joke since the beginning of hawken

you can try to explain as much you want

 

it is not working

[CraftyDus]

He sed cumlog.... heh heh

[kasei]

Luna


Me


Sorry, I'm off work, I limit my CPU on not peak hours.

Edit:

I assume it's nice work though, good job.

Edited by kasei, 27 May 2015 - 04:46 PM.

[Draigun]

why am i laughing ???

 

the matchmaking mmr and all that stuff was a joke since the beginning of hawken

you can try to explain as much you want

 

it is not working

 

Your ignorance to the power of mathematics and statistics is, if anything, frightening. Although given the fact that most of what has been described never sees the light of day to gamers around the world, I can't blame you. It's something developers generally don't want to delve into with their communities.

 

Imagine this: Hawken was developed with a population base so large that it would be in the millions. Those were the goals for previous publisher Meteor Entertainment.

 

In terms of how realistic it can get, a larger player base will allow the conceived system to perform far more effectively, as compared to a lower player base. Increasing our player base will also increase the effectiveness of the systems used to place you into matches. Additionally, a larger span of players will increase the effectiveness of auto-balance. In layman's terms, chances of unbalanced matches go down, while statistical outliers in MMR also go down.

 

Excellent work on breaking it down, OP. I'm not so sure about some of the flags though, I couldn't imagine the formula having a limit, and for what purpose.

[plons]

 

 

Your ignorance to the power of mathematics and statistics is, if anything, frightening. Although given the fact that most of what has been described never sees the light of day to gamers around the world, I can't blame you. It's something developers generally don't want to delve into with their communities.

 

 

 

 

 

 

mathematicians have one big problem when they come with stuff like that .......they always forget the factor human

 

mmr does not work because human is a to big variable .......

 

once you can convert human into a math formula  this mmr thing will work ....

 

 

 

in the whole mmr explanation you don't see the variable human

 

 

 

i have deep respect for mathematicians because without them we should have no computers and electronics and much other stuff

 

 

 

 

 

 

 

 

Edited by plons, 28 May 2015 - 03:09 AM.

[StubbornPuppet]

I think that one, simple, adjustment to the equation would really help with how MMR is calculated and used in the matchmaker and team balancer:

Remove any players who have not logged in during the last 90 days... maybe 60 days.  The thing is, probably 90% of the players in the overall game have not played more than a minimal number of hours and are long, long gone.  Most of these "long gone" players probably did very poorly at the game too - which makes sense, because we don't like to keep doing things we aren't good at.  With such a vast majority of the "player base" that the current MMR calculation is counting being irrelevant, how can it possibly do a good job of giving a realistic representation of what is actually happening among active players?

[Badtings]

Thanks for all your work, OP!

[n3onfx]

mathematicians have one big problem when they come with stuff like that .......they always forget the factor human

 

mmr does not work because human is a to big variable .......

 

once you can convert human into a math formula  this mmr thing will work ....

 

 

 

in the whole mmr explanation you don't see the variable human

 

 

 

i have deep respect for mathematicians because without them we should have no computers and electronics and much other stuff

 

 

 

 

 

 

 

 

 

 

The "variable human" doesn't mean anything in this context. You're just rambling about a random pseudo-philosophical buzzword in a situation where it has nothing to do.

[plons]

The "variable human" doesn't mean anything in this context. You're just rambling about a random pseudo-philosophical buzzword in a situation where it has nothing to do.

so players are not human ????

[Hyginos]

Mathematicians have one big problem when they come with stuff like that: they always forget the factor human.

 

The entire system described in the OP, the MMR system, and a HUGE portion of the field of statistics is devoted solely to understanding the human factor. Please also limit use of punctuation and carriage return to single quantities.

 

 

I think that one, simple, adjustment to the equation would really help with how MMR is calculated and used in the matchmaker and team balancer:

Remove any players who have not logged in during the last 90 days... maybe 60 days.  The thing is, probably 90% of the players in the overall game have not played more than a minimal number of hours and are long, long gone.  Most of these "long gone" players probably did very poorly at the game too - which makes sense, because we don't like to keep doing things we aren't good at.  With such a vast majority of the "player base" that the current MMR calculation is counting being irrelevant, how can it possibly do a good job of giving a realistic representation of what is actually happening among active players?

 

Maybe I misunderstood the OP, but I'm pretty sure any players MMR will only change based on the population of the server he is in and his score therein. That means that inactive players do not affect the MMRs of active players at all. A mass of inactive players should only show up on a distribution of all players, which may be nice for people to look at but does not impact matchmaking.

 

@OP: IIRC the Gliko system has an uncertainty variable attached to every player that changes based on time inactive, recent record, and/or similar stuff. Do you think the Hawken implementation includes that?

[TURDxSANDWICH]

why am i laughing ???

 

the matchmaking mmr and all that stuff was a joke since the beginning of hawken

you can try to explain as much you want

 

it is not working

 

 

You appear to have not understood the OP.  There is no attempt to dissect, explain, defend, or critique the Matchmaking process.  This is only a fairly clear and robust explanation of how MMR is calculated and how it is affected by a player's performance relative to the average MMR on that server in each match.

 

Nice work Ishmael!  I really enjoyed the slides and this will be great for future reference about what MMR is.  And hopefully, once a better understanding is reached among players and devs about what MMR is, then perhaps it can be used more effectively in the matchmaking process.

[TheButtSatisfier]

mathematicians have one big problem when they come with stuff like that .......they always forget the factor human

 

mmr does not work because human is a to big variable .......

 

No. Humans are mathematically predictable in aggregate. There's nothing mysterious about a human's behavior that makes mathematicians/statisticians throw up their hands and say, "Damn, if only statistics could apply to people! It's so frustrating, their free will makes them impossible to predict to any useful degree. I'd better stop doing statistical analysis to determine the impact of advertising or perform election predictions because I keep forgetting that unquantifiable human factor people keep telling me about."

 

If you were right then Google couldn't do traffic prediction, entire economic models would be useless, and everything we apply big data to would fall apart because they are based on how humans behave in those contexts. Those fields/applications wouldn't exist or they would be pseudoscience at best.

 

But they do exist, and they do work, so you're wrong.

Edited by TheButtSatisfier, 28 May 2015 - 10:15 AM.

[TURDxSANDWICH]

so players are not human ????

 

Again, there appears to be a misunderstanding here.  The MMR calculation is precisely an attempt to quantify human interaction within a match of Hawken taking into account players' relative skill.

[zorin1]

I guess the questions really is what formula should be used.  I have noticed that my MMR takes a big dive when I am in a 24 Man match if I don't place high to the top.  Mostly if you get in during the middle of a match.  Maybe your MMR should not be calculated if you are not in the match for the whole time.  One other thing that I have noticed is if you leave before the match is over that your MMR is not affected.  So if you are in a Siege game for like 15 mins and your team is getting killed and you are at the bottom if you leave before the match is over, your MMR stays the same.  

 

I do think this explains why my MMR drops from a 12 man team matches.  If you join a low MMR game and only play for like 5 mins and you do not move up then your MMR takes a big hit.  I think in one match that my MMR dropped 20 points.  And it is a lot harder to get those 20 points back then it is to lose.

[StubbornPuppet]

Maybe I misunderstood the OP, but I'm pretty sure any players MMR will only change based on the population of the server he is in and his score therein. That means that inactive players do not affect the MMRs of active players at all.

 

Sort of, but the point I'm making is that the overall basis for the MMR calculation is still averaging using ALL players, active or not.  I'm not talking about how it caculates how much MMR score a player gains or loses during a match.  So, a players mmr score is based on numbers that include inactive players, but how much it goes up or down in a match does is not.

 

This is why the scales look so enormously tipped towards the 1200mmr player base.  If any player who has not been active in the last 90 days were taken out of the calculation, I guarantee that everyone's actual mmr would change drastically.

Edited by StubbornPuppet, 28 May 2015 - 11:26 AM.

[Call_Me_Ishmael]

I guess the questions really is what formula should be used.  I have noticed that my MMR takes a big dive when I am in a 24 Man match if I don't place high to the top.  Mostly if you get in during the middle of a match.  Maybe your MMR should not be calculated if you are not in the match for the whole time.  One other thing that I have noticed is if you leave before the match is over that your MMR is not affected.  So if you are in a Siege game for like 15 mins and your team is getting killed and you are at the bottom if you leave before the match is over, your MMR stays the same.  

 

I do think this explains why my MMR drops from a 12 man team matches.  If you join a low MMR game and only play for like 5 mins and you do not move up then your MMR takes a big hit.  I think in one match that my MMR dropped 20 points.  And it is a lot harder to get those 20 points back then it is to lose.

 

35/24 = 1.45

 

If you finish 13th, your score will go down 1.45 x your log curve score (which depends on your MMR and the server MMR).  You can finish 5th (of 12) on your team and still be 13th overall in the server.

[Call_Me_Ishmael]

@OP: IIRC the Gliko system has an uncertainty variable attached to every player that changes based on time inactive, recent record, and/or similar stuff. Do you think the Hawken implementation includes that?

 

I think so.  I didn't add it to the slides, but I can when I have some time, thanks.

[Call_Me_Ishmael]

Maybe I misunderstood the OP, but I'm pretty sure any players MMR will only change based on the population of the server he is in and his score therein. That means that inactive players do not affect the MMRs of active players at all. A mass of inactive players should only show up on a distribution of all players, which may be nice for people to look at but does not impact matchmaking.

 

 

Yes, this.

[hoghead]

Crafty your a goof !!!!

[MechFighter5e3bf9]

only ashfire should know  the secrets of mmr and no one should be abusing any weaknesses to raise theirs, long as its as unpredictable as it is many players would not dedicate much effort in a style they dont enjoy trying to climb a stat and just join 3 star servers, or perhaps 1 star if they are particularly handicap or just not tryin hard playin relaxed