1. Hello and welcome to the brand new home for PlayDota!
    Please read through our Welcome thread to see what's new!
    Dismiss Notice

Improving the MMR calculation

Discussion in 'Visual & Gameplay Ideas' started by a_lost_hero, Nov 27, 2016.

  1. a_lost_hero

    a_lost_hero New Member

    Nov 27, 2016

    I've recently watched Purge's great series on MMR, over at https://www.youtube.com/watch?v=MVZdeTGwJhU
    If you haven't already, i highly recommend watching them all!

    Now, during the 3rd and 4th episode the idea struck me... why isn't your 'effective MMR' taken into consideration when adjusting for a win/loss?

    Why would we want to mess with the MMR system, i hear you cry. 'Get good, you suck' is the common intelligent level of debate.

    Well, believe it or not, many players feel that something is missing/lacking with the MMR system... and I believe Purge has the answer!

    According to his videos, MMR is calculated thusly:
    y = Team Average MMR compared to enemy team
    V = -y
    0.06xV +/- 25 (Gain/Lose MMR)

    Essentially this means that if the MMR between the two teams is Zero, then you will lose or gain 25 MMR.

    All well and good so far? Well, no, not really. When I first read the MMR changes when it was first introduced, it promised me individual performance adjustments and greater MMR changes if my team was considered less likely to win, right?

    Well, down in the 'scrub' brackets, there are so many players that it is really easy for the matchmaking system to find two teams with the same MMR... so you're stuck with 25 MMR increments forever.

    What's the problem with that? I hear you say. Well, what happens if 3 players on your team random their heroes? Or you pick last and are forced to pick a position or counter pick that you're not familiar with? The current calculation doesn't take this and many other factors into consideration.

    So, I bring to you, the new and improved MMR calculation:
    0.06xV +/- 0.06xE +/- 25

    What is the E? The E is Purge's 'Effective MMR'.
    This is your effective MMR on the hero you've ended up picking:
    (Win Rate on Hero - 50%)/100 + 1 all times by your MMR.
    So if your win rate is 56% and your MMR is 5,000, this should equal an effective MMR of 5,300.
    If you add up each teams effective MMR and calculate the difference... this is E.

    So now you have your real MMR difference, which is inevitably Zero, and your Effective MMR difference, which could add that wild and much needed calibration and adjustment factor that the MMR system is sorely lacking.
    What might have been a 25 MMR increment is now more variable depending on how likely it was you were to win based on the actual heroes you picked.

    Accounts for the actual game you're playing, not just your average MMR over all previous games.
    Accounts for hero spammers, randoming, playing heroes out of your comfort zone.
    Someone gets to run some tests with big data (sounds fun to me, so it's a benefit!)

    Valve has to actually be reading this post.
    People will prove they won't understand what i'm saying by spamming in the comments and detract from serious discussion.
    It won't account for players abandoning the match or sitting in the fountain doing nothing (but what can?).

    Thank you for reading, hope it made sense, ask for clarifications below :)

    EDIT: Essentially, if I could get my hands on the relevant big data of all previous matches since MMR was introduced, I'd love to re-run the calculations to see what people's MMR would actually be today. I'd expect a lot of movers and changers.
    Last edited: Nov 27, 2016
  2. dewouter

    dewouter Moderator Staff Member

    Oct 27, 2009
    It's pretty hard to accurately check the results of this without checking the data, but I'm pretty sure there is more than a small chance that the sum of all points won't be 0. That is going to skew any rating system. Honestly it's all about winning and nothing else really matters, I don't get why people want to punish spammers more.

    It might be interesting to see some results nevertheless
  3. mapdesigner

    mapdesigner Member

    Aug 11, 2011
    its only 4 episodes I watched it all. I thought more is coming...