The tl;dr
- The current matchmaker frequently chooses teams with a marked disparity between the MMR of the Damage and the Support players.
- There is a specific mathematical reason to believe the matchmaker would behave this way, and that explanation does not require detailed knowledge of the matchmaking algorithm.
- The down-ranking MMR behavior leads to mismatched teams, which in turn lead to bad games for everyone.
- The down-ranking MMR behavior is probably an emergent property of the matchmaking system itself, and not a deliberate ploy by Blizzard.
- The result is still deeply frustrating, especially to Support mains.
The Introduction: “I need healing!”
If you solo queued for Support during Overwatch Competitive Season 2, your matches were probably bad – overwhelmingly bad. It felt like you were slotted into teams that could not function, teams that stumbled through an entire match only to hand it all to your opponents in some completely unnecessary, avoidable way. The losses came in punishing streaks, and even when the streak broke, many of the wins felt less like clean accomplishments and more like landing on the winning side of a mismatched blowout.
Let’s be realistic though: You lost at least some of those games through no one’s fault but your own. We all make mistakes, and we all have room to improve. It would be unproductive, and untrue, to believe otherwise. But let me give you the benefit of the doubt, and assume that you can tell the difference between a losing game and a bad game. Even a win can make a bad game, and even a loss can make a good one. But when wins and losses both feel disjointed and inexplicable, that’s a clue that something is really wrong, and it’s not just something about you. If it would be unreasonable to attribute every single loss to pathological matchmaking or bad teammates, it would be just as unreasonable to claim that every single player (and there are a lot of them) who has joined the outcry against the Bad Games of Overwatch Competitive Season 2 is somehow just making delusional excuses for their own deficient skill. The games have been bad – really bad – and something is going on here.
Or is it? Many folks in the community would have that anyone struggling through losing streaks is simply being put in their proper place by the matchmaker. The Developer Blog acknowledged problems with matchmaking in its two-part post (see Part 2 especially), but was vague in explaining the origin and mechanism of the problem, and offered little help to players trying to separate their own failures from those of the system. The result has been a toxic mix of confusion, blame, and superstition. Is the matchmaker somehow forcing a 50 percent win rate? Are players being segregated into secret queues of triumph or doom, without their knowledge? Are Supports all just inveterate whiners who can’t accept losing? Is the whole thing part of an insidious scheme to ramp up player engagement and move more in-game purchases? The arguments rage all season, but all of the arguments are unwinnable because they are based on nothing more than opinions and anecdotes.
In the interest of correcting this state of affairs, I present you with something I have not seen yet amid all this mess: A logical argument that could be falsified by analytical argument or the right data. It may or may not be correct, but even if it turns out to be incorrect, its negation still tells us something we did not know before. The thrust of the argument is this:
Your games were bad because there were not enough Support players to make decent matches in a timely fashion, so the matchmaker traded speed for quality in the interest of keeping the peace.
I made this assertion as part of a long reply on Part 2 of the Dev Blog’s matchmaking series, but I didn’t go into detail because my reply was already quite long. Moreover, my claim isn’t self-evident, so I would like to expand upon the justification a little further. My hope is that:
- Other Support mains who suffered through Season 2 might find their struggles vindicated.
- Blizzard might affirm what I have suspected, or show evidence that refutes my claim.
- We can look ahead to what might or might not improve for matchmaking in Season 3.
The Argument
Premise 0: The matchmaker aggregates individual player MMRs into a single score when attempting to choose teams.
The Dev Blog writes:
[W]e’re implementing some updates to the matchmaker that try to place pairs of players with similar MMR on each role on either team. The goal with this change is to make the average MMR between each role more evenly matched to each other instead of looking more broadly across the entire team to balance things out.
It follows logically that there would be no need to change the matchmaking algorithm in this way if matches were already aligned by both MMR and role. Likewise, it seems implausible that the matchmaker is aligning individual players by MMR; if this were the case, there would be no need to “[look] more broadly across the entire team to balance things out”, since the matchmaker could simply choose players in opposing pairs.
When I say that the “matchmaker aggregates individual player MMRs into a single score”, it might be helpful to think of this “aggregate” as the mean average of the MMRs on the proposed team. The actual matchmaking algorithm probably calculates a more sophisticated function than a simple average, and while we don’t know the definition of that function (unless, perhaps, we work for Blizzard), we can conclude that it must combine the MMRs of individual players in some way that would allow it to compare one possible team to another and to conclude that one team was “stronger” than the other. So, think of this function as the team-average MMR, with the asterisk that the real function is probably different.
Premise 1: There are far more players queueing for Damage than for Support.
Hopefully this fact is already self-evident to anyone who bothers to look at the queue times. I have never seen Competitive queue times for Damage listed as better than “<4 minutes”, and at various times in the season, that threshold has been much higher. By contrast, the queue for Support is seldom worse than “<2 minutes” – and is usually less. Since each Role Queue team requires two Damage and two Support, there are the same number of available matches for both Damage and Support. If Damage players have to wait longer, it must be that there are more of them waiting for a match.
There seems to be general agreement in the community that Damage players outnumber Support by quite a bit. I am not aware of any numbers that show the magnitude of the discrepancy, but its seriousness is further corroborated by the broad attention it has received, both from players and even from Blizzard.
Premise 2: Given a randomly selected pool of players, there are more ways to form an imbalanced team than there are ways to form a balanced team.
In my original reply on the Dev blog, I glossed over this point as a “straightforward combinatorial argument”. While a mathematical proof would be too much for this forum, the point still deserves some justification.
Imagine you have a well-shuffled deck of playing cards, from which you draw four cards. (Let’s assume we remove the jokers.) How many different ways could you draw a hand consisting of only a single face value, i.e. every card is a 2, or a 3, or so on? Since the order of the cards doesn’t matter, you can draw as many hands as there are face values: there is one hand for the 2s, one for the 3s, all the way up to the kings and the aces, which would make for 13 possibilities altogether.
By contrast: How many possible four-card hands could you draw from the deck altogether, if there were no restrictions on what those cards might be? If you remember a little bit of math from school, you might work this out and realize the number is much, much bigger than 13 – like, thousands of times bigger. (For the record, the answer is: 270,725 possible four-card hands.)
In this example, the cards are our Overwatch players, and the face values of the cards are the MMRs of those players. In normal Overwatch Role Queue, we are choosing five players instead of four (at least, that’s how it works as of this writing), but the disparity between the number of balanced and imbalanced teams remains – in fact, the disparity gets bigger. There are simply more ways to form an imbalanced team.
Premise 3: Matchmaking must trade off between match quality and wait time.
This fact is fundamental to the problem of matchmaking generally, whether in Overwatch or elsewhere. Given a queue of possibilities and a player to match, the matchmaker has to choose between selecting a “good enough” match, or waiting for new players to enter the queue who might make for better matches. There is no single matching strategy that can optimize both match quality and wait time, since it is almost always possible that a better match could appear in the future, and since there is always a point at which wait times become unacceptably long.
This trade-off has an important consequence: in order to decrease wait times, you must be willing to accept lower-quality matches. If you loosen restrictions on which matches are “good enough”, then more matches become possible, and there is less need to wait for new players to enter the queue.
Conclusion: In order to mitigate queue times, Support players are frequently matched to Damage players whose MMR is lower than their own.
Let’s unpack how the premises lead to this conclusion:
Since there are more Damage players than Support (Premise 1), there is a stronger pressure to match a given Damage player to a game (Premise 3) than there would be for a Support, since Damage players will, on average, have to wait longer. When forming a full, two-team match, the matchmaker must only select two teams whose respective team-average MMRs (Premise 0) are not too different. In particular, the match need not have a particularly high or particularly low average MMR – the only sure requirement is that the team-average MMRs should not be too far apart. Since there are more ways to form an imbalanced team (Premise 2), it will be easier to match a given Support with a lower-ranking Damage than with a similar-ranking Damage, which means such matches will occur more often.
Consequences and Speculations
Aligning MMR by role, as proposed for Season 3, could mitigate this problem. Although I would conjecture that Supports could still be teamed with mismatched Damage players at a comparable rate, the opposing team would have a similar composition on average.
But queue times may increase. Since the matchmaker will have to be more choosy about its matches, it will have fewer acceptable matches to choose from.
Without data, it is hard to say how strong this effect might be. That said, I would expect the severity of this problem to scale with the magnitude of the disparity between the Damage and the Support queues. That is, the problem should get worse as the Damage queue gets longer relative to the Support queue.
This same argument might suggest high-MMR Damage players are matched with lower-MMR Supports. And I see no evidence against claiming as much. However, I have seen far more Support mains complaining about their Damage teammates than the other way around. This asymmetry raises some interesting questions about why Support players might feel the disparity worse, and some of these might already be hinted at in the broader discourse. If we had data on the outcomes of imbalanced teams, we could conclude whether Damage players do, in fact, have greater capacity lift match outcome than Supports. It’s clear that many players suspect as much.
It would be interesting to see if this phenomenon is concentrated in certain regions of the rating hierarchy. In particular, in Part 2 of the matchmaking explainer, the Dev Blog notes that:
[W]e have data showing that matches become less predictable the lower the average rating of a match.
This fact has potentially far-reaching consequences, both for this argument and for many others.
If the troubles are not uniformly distributed across ratings, it could explain why so many higher-rated players claim not to see the problem, while lower-rated players languish. It might not be a simple case of personal bias – those two groups might actually be sampling their games from markedly different distributions, and thus really observing different things.
The Competitive queues are an especially complex feedback system, and it is unclear how that system’s stability, or lack thereof, might play a part. Many commentators note that wins and losses seemed to run in streaks throughout Season 2, and this kind of oscillatory behavior is exactly what one might expect from a statistical estimator that is struggling and failing to converge. Because there is no “ground truth” MMR – i.e. MMR can only be defined in terms of other players, and “discovered” by playing matches which are, in turn, chosen based on MMR – there is a lot of room for chaotic behavior. That behavior could play a part in this phenomenon, or could be a wholly orthogonal problem of its own.
This explanation is not complete. In no small part because we do not know exactly (or even approximately) how the matchmaker uses MMR, and because we do not have public match data at any scale that would affirm or refute particular claims about patterns in game outcome.
In Summary
If you solo queued Support in Season 2, there is a strong argument that your games really were as bad as you thought, and that matchmaking played at least some part in creating that problem. Even without data on match outcome or knowledge of Blizzard’s algorithm, we can use straightforward observations about the game to draw some logical conclusions. Those observations can in turn explain how dynamics of the queuing system could bring about your bad season, and could do so without the need to blame yourself for every loss or to posit some algorithmic conspiracy by Blizzard.
It was a bad season, yes, but there is at least an understandable reason. I hope your next one goes better.
\0
