Page 1 of 2 1 2 LastLast
Results 1 to 10 of 14

Thread: Matchmaking Explained

Hybrid View

Previous Post Previous Post   Next Post Next Post
  1. #1

    Post Matchmaking Explained

    Surprisingly not a suggestion or complaint thread. This is a short playdota article I wrote to explain matchmaking systems.


    I won't pretend to be an expert, but I have an understanding of matchmaking algorithms. If anyone would like to correct any of my points, feel free to do so.





    Your Matchmaking Score
    -----------------------------



    Actual matchmaking has nothing to do with "what kills you have" as that is a factor that isn't entirely affected by skill, and has negative results on players who aren't selfish. Which is why such an algorithm would actually be flawed.

    Matchmaking algorithms are based on probabilities. If you are playing against players of an equal matchmaking score, you should have a probability of winning around 50%. The ideal of all matchmaking algorithms is to have you around a 50% win rate, because you should be placed against equally skilled players all the time.

    On average, you out perform players below your current skill level, and perform worse than those above. Though you often will win against greater players and lose against worse players, as this is only an estimate of what "should" happen.

    If you win against someone you should win against, your rating stays around the same, but if you beat someone above you, your rating will increase. The opposite is true if you lose.

    The difference the loss or win makes is actually calculated by the difference between your skill levels. Beating someone way over your skill level changes your ranking much more, than the very tiny boost you'll get for beating someone only slightly over your skill level.





    Their is an additional variable that increases the magnitude of how much you lose or gain each game, which is used to make the first several games for new players worth much more toward their ranking than any other game, because it needs to get them somewhere in the system (though it's not an accurate ranking quite yet, this speeds up the process to approach it). This variable tends to ease out and eventually stop increasing the magnitude of each change when you have lots of matches played.

    These matchmaking algorithms also tend to take how often you swerve away from your average probability of winning as a factor of how much your matchmaking score will change, to avoid huge jumps for players who have lots of wins against better players and lots of losses against players who are worse.

    If we are talking about what determines your performance, it generally will be just how much you have won vs lost, and in a team game it will likely sum up the two teams player scores to determines the probability of who should win, and what impact that will have.





    Many systems take into account how close each game was, though we do not know if this is a factor used in Dota 2's matchmaking system.

    Some players like to party with friends of much lower level than they are. This is why you sometimes get players who are a lot lower level than you are in the game, and though the system won't jump this lower level player for high level wins if it already understands that players matchmaking rank, it could mess up if they are still in their first few games.

    Smurf detection is it's own complicated issue, but we know for sure that the system checks if you completely pub stomped almost every one of your first several games, and if you did, it will skyrocket your matchmaking ranking. So high level players with new accounts will quickly leave the low level matchmaking pool. This is the only scenario where kills and deaths have any effect on your matchmaking rank.

    The system can't grantee the outcome of a game even with what is supposed to be a very balanced match. Also, we have to consider that dota 2 games are also affected by picks, communication, and who is playing what hero, which is even MORE of a nightmare to judge, in comparison to determining the skill difference when both players are playing the same heroes/soldiers/have the same units.






    Finding a Match
    -------------------



    When you enter matchmaking, you enter a localized pool of eligible players to join a game.

    What pool you are in is determined by the region you selected (for example, US East, Europe West, etc...). If you selected multiple regions, you are in every pool you selected.

    This pool lists you by your matchmaking ranking, and during the time you wait, it's going through players within a certain range of your skill level and placing them in a match lobby with you.





    If it doesn't find enough players within that range the first time, then it will repeat the loop, except with a less restrictive range, meaning their is a larger skill differences between you and these players, than the difference between you and the players it has already found.







    This loop repeats until it finds ten suitable players, and has balanced the teams between those ten suitable players found. Once you are in a match, you leave your pool(s) and connect to the game server.

    So the skill difference between you and the players in your game is determined by what players are available, and even if their is a large skill difference, it's still better than you waiting in the matchmaking pool for hours, waiting for that perfect match.

    Specific to Dota 2, the larger your party size going into a match, the lower the range of skill between you and your opponents it will tolerate. So you will wait a little longer for what will be likely be a better quality match.

    So you are more likely to find a good match if you have a ranking with a lot of players, which is around 3500, at the time when the most players are on, and in the highest traffic region.

    So unfortunately, this means that in the reverse scenario, if you are near the highest or lowest ranking possible, where there are fewer players, you play around 3-5am, and if you live in a region with the least players, you will have matches of lower quality, and longer wait times.

    This is why some players decide to play in other regions than their own, because they want to avoid the tremendous wait times they get in their region. Some regions also have terrible internet, so they will get the same ping nearby as they would in another region anyway.






    Potential Inaccuracies
    -------------------------



    There is only one issue I have gathered that has strong evidence for it, to following is a direction quote from the playdota member xpforever.


    Quote Originally Posted by xpforever View Post
    This is not a complaint thread, this is a partially informational and partially speculative thread that will attempt to explain why you have so many problems with the matchmaking, and why players from some regions seem to cause problems in your games.

    This is a problem with how rating systems work in general. Suppose we have one big happy dota region. The skill distribution and MMR would be consistent, and would look like something like a bell curve (note that the numbers aren't the same scale that dota 2 uses):


    In fact, if we took a sample of players, gave them their own Dota 2 region, and had them play lots of matches against each other (but never against people outside of this region), then the MMR curve would look more or less like that, with the median being 1500 (or whatever Valve uses). It doesn't matter if you took a bunch of completely new players or a bunch of professional players. The average MMR would be the same since it's relative to the group.

    Now, what happens when we add multiple regions? The problem is that the average skill level of every region won't be the same (no offense meant to anyone). There are highly skilled Russian players, and there are low-skilled US players, but just assume that the average skill level of the Russian region will be lower. If the regions were completely separate (region-locked), we would end up with a distribution that looks like this (note that this is highly exaggerated):


    Now in theory, if nobody played cross-region, this wouldn't be a problem. MMR would generally be accurate within regions; an average NA player would have 1500 MMR, while an average RU player would have 1500. Despite there being a disparity in terms of actual skill, this wouldn't be a problem since they would be playing within their own region. NA players would play with equally skilled NA players, and RU players would play with equally skilled RU players. And, as shown with the first graph, if there was significantly more cross-region play, it would fix the matchmaking since it would push everyone towards their true MMR relative to the world rather than just their region.

    The problem comes with Dota 2's current matchmaking situation where MOST matches are played on one's originating server, but a few are played cross-region. This means that in order to get evenly-skilled cross-region matches, you need to have world-relative MMR. However, since a significant amount of matches are played in-region, MMR will be skewed towards region-relative MMR.


    Does Valve know about this?

    Yes, they acknowledged it as a problem on the dev forums a while back.
    So our matchmaking ranking is not a true world relative ranking, which may cause slight inaccuracies for cross-region games.










    Sources:

    Official explanation by Valve about the Dota 2 matchmaking system
    Description of the True Skill algorithm that Xbox live uses
    Valve Employee Comments on Matchmaking
    Dota 2 General Information About Matchmaking
    Wikipedia "Elo Rating System"
    Potential inaccuracy thread by xpforever"
    DotaMetrics Unofficial Matchmaking FAQ - Part 1
    DotaMetrics Unofficial Matchmaking FAQ - Part 2
    DotaMetrics Unofficial Matchmaking FAQ - Part 3
    League of Legends Matchmaking Explained
    Last edited by Burning Titan; 10-22-2014 at 01:57 PM.

  2. #2
    Basic Member Concede's Avatar
    Join Date
    Nov 2013
    Posts
    294
    the first part is not accurate during calibration and/or your first games (if you got to level 13 without playing non-ranked matching making)

    Once calibration is over, what you are saying is mostly correct.

  3. #3
    Basic Member SabotaZ's Avatar
    Join Date
    Mar 2013
    Posts
    1,061
    Quote Originally Posted by Concede View Post
    the first part is not accurate during calibration and/or your first games (if you got to level 13 without playing non-ranked matching making)

    Once calibration is over, what you are saying is mostly correct.
    this. before the introduction of ranked mm, I remember a guy with 6 games playing in the 3rd top live game. Reason he was there is he had stomped on his previous games with ~40 kills/game and he got skyrocketed. Quite unfortunate though that he never belonged there and he ruined the game for his allies.

  4. #4
    You'll notice that I actually have a short paragraph on smurf detection where I talk about that.

    The opening paragraph was for the general audience.

  5. #5
    Basic Member
    Join Date
    Apr 2012
    Posts
    152
    How does it place me if my (party) MMR is still under "calibrating" if I queue in a party? For example, my solo MMR is 3.9k, and yesterday I queued with 2 friends with 2.5k/2.9k MMR (both solo and party), and I noticed that the opponents' skill level was pretty low.

  6. #6
    Basic Member
    Join Date
    Feb 2012
    Posts
    618
    Quote Originally Posted by Burning Titan View Post
    There is only one known complaint about valve's matchmaking with strong evidence.
    The following is directly quote from a post by xpforever at playdota.


    Quote Originally Posted by xpforever

    This is a partially informational and partially speculative thread that will attempt to explain why you have so many problems with the matchmaking, and why players from some regions seem to cause problems in your games.

    This is a problem with how rating systems work in general. Suppose we have one big happy dota region. The skill distribution and MMR would be consistent, and would look like something like a bell curve (note that the numbers aren't the same scale that dota 2 uses):


    In fact, if we took a sample of players, gave them their own Dota 2 region, and had them play lots of matches against eachother (but never against people outside of this region), then the MMR curve would look more or less like that, with the median being 1500 (or whatever Valve uses). It doesn't matter if you took a bunch of completely new players or a bunch of professional players. The average MMR would be the same since it's relative to the group.

    Now, what happens when we add multiple regions? The problem is that the average skill level of every region won't be the same (no offense meant to anyone). There are highly skilled Russian players, and there are low-skilled US players, but just assume that the average skill level of the Russian region will be lower. If the regions were completely separate (region-locked), we would end up with a distribution that looks like this (note that this is highly exaggerated):


    Now, if nobody played cross-region, this wouldn't be a problem. MMR would generally be accurate within regions; an average NA player would have 1500 MMR, while an average RU player would have 1500. Despite there being a disparity in terms of actual skill, this wouldn't be a problem since they would be playing within their own region. In theory, NA players would play with equally skilled NA players, and RU players would play with equally skilled RU players. And, as shown with the first graph, if there was significantly more cross-region play, it would also fix the matchmaking since it would push everyone towards their true MMR relative to the world rather than just their region.

    The problem comes with Dota 2's current matchmaking situation where MOST matches are played on one's originating server, but a few are played cross-region. This means that in order to get evenly-skilled cross-region matches, you need to have world-relative MMR. However, since a significant amount of matches are played in-region, MMR will be skewed towards region-relative MMR.

    Does Valve know about this?

    Yes, they acknowledged it as a problem on the dev forums a while back.
    This and the fact that it takes way too long to go up or down on the ladder are the only real issues with matchmaking, that are actually fixable without needing a supercomputer to account for a seemingly infinite amount of variables. If that issue and the one I'm pointing out were fixed, then people would literally have zero excuse about how bad or good matchmaking is, and would instead have to bitch about the nature of pubs, rather than the system that creates them

    Edit: Then again, it means Valve has to hurry the fuck up and provide servers for the colossal amount of players that don't have any yet, look at South America, the only server it has is for Brazil and only allows brazilians in it due to ping, on top of them being the only country in whole South America that speaks Portuguese, of course

    It baffles me when people wonder why LoL is so popular, well no shit it's popular in some places, it paid more attention to the playerbase from these places than Dota2 did, LoL has really good south american servers, there's a reason they aren't considered cancer over there, or at least not as much, while Dota2 is described as "The game that made me xenophobic" or "The game that taught me profanities in a different language, and how to hate everybody that speaks in that language"
    Last edited by Infer; 03-18-2014 at 03:30 AM.

  7. #7
    Lucid and useful. It even has sources. Thank you for sticking this up, BT.

    Sideways a bit:
    I also still suspect they sort people by use of voice chat too. I always tend to run into those people in clumps.

  8. #8
    Basic Member
    Join Date
    Sep 2013
    Posts
    41
    Lets add into the equation, feeders, leavers, afk ragers, carry supports, terrible captains (5 supports/carries etc...), boosters, arguing teams etc...
    And welcome to the mmr trench.

  9. #9
    Fixed a few typos.

  10. #10
    Basic Member J2Krauser's Avatar
    Join Date
    Sep 2013
    Location
    Mother Russia
    Posts
    1,028
    Why did you have to bump this?

Tags for this Thread

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •