Unless someone can post a link, I don’t believe they’d have such huge divisions. They have literally 5000 increments for SR, but only 13 for what they actually match you on? Blizzard is dumb, but I don’t think they’re THAT dumb…although it would explain a lot about the quality of matches.
Actually as writing this, I do recall a Jeff/Seagull interview where he mentioned it goes “a little above 3” so I believe the 0.5 increments even less now.
Inflammatory comment
plus 3 to minus 3, in 0.5 increments would equate to 12 divisions. so 5000/12 would be roughly 416sr diff per bracket, which is about equivalent to a rank, since a rank is 500sr diff
Clumping everyone who is 2500-2916 is pretty horrible for matchmaking. It’s a good thing they don’t use the .5 as people are claiming.
It’s clumping but with a 500 sr swing. So it depends on the highest sr person in the group.
ALSO we don’t know the percent win. They might view that as 50/50 still.
I had a game it was bronze silver gold, bronze silver gold. (Just happened to remember the pattern) vs bronze silver gold silver silver gold.
Very bizarre stuff going on. And your telling me there’s not enough gold nor silver nor bronze players to make up one team?
This is why that team average nonsense has got to go. Either each player in each role… has to face someone relatively equal (silver DPS must match a silver DPS) or keep everybody on each team the same rank (all silver.) It could even quell people complaining about MMR (me being one of them)
Would be interesting (if possible) to see if matching individual to individual would yield interesting results. Like matching one support on one team to one on the other and not considering averages as much since each player should theoretically have an equal on the other team assuming they’re accurately rated.
Not necessarily advocating for it as I can see some flaws but just a thought experiment.
I am not advocating truly for it, I’d prefer a role average rather than a team average. Because I do generally think that each has their own “average” player, and due to small tank base… vs. the large DPS base - an average player wouldn’t necessarily be the same. Obviously the ends of the bell curves would be different , they are the best of the best, and the worst of the worst performers. But I guess that currently the idea of average team SR numbers are very off due to it originally based off of one number (Before Role Que) instead of 3 (After Role Que.) and I think the middle ranks pay the most price for it.
Thats the most logical explanation, right?
But then where you will get 6 pairs of players with exactly same MMR, hence chances are not equal to some degree, and those never will be equal. Cause one of players likely to show up drunk, 2nd is upset cause wife dropped him etc
what’s this argument? It doesn’t make sense.
its called Bayesian probability. the algo takes statistical data to calculate the probability of a match outcome.
It doesn’t mention Bayesian probabilities.
why would they, its the actual mathematics behind the matchmaking. the average player is not going to understand so they just use the term mmr.
So your going deeper than mmr?
yeah. their matchmaker prob looks like a big pile of algebra and the “mmr” bit will be just one part. but essentially what it is all doing is taking data variables and using baysian probability to quantify a prediction of the match outcome.
how i think the matchmaking and mmr works is the game will start out by assuming the player is of average skill since probability wise that is the best guess since statistically most people will be average.
so it assumes you are average and puts you in average games and depending on the outcomes and performance it updates its estimate of your skill level and how confident it is in that estimate, which is what Bayesian probability enables it to do.
the mmr component allows it to basically remember your peak performance and factor that into its confidence of your estimated skill rating and thus how you gain or loose sr.
so for example say it put you in gold and you went on a decent run of wins and got to plat and even won games in plat but then you hit a loss streak and fell down to silver.
although you are now in silver the matchmaker wouldnt be very confident that you are a silver player since it has data that you were able to win games in plat. where as if you started in gold and immediately fell to silver it wont have much confidence that you are plat or higher.
Bayesian probability allows the matchmaker to be constantly adjusting its estimate of your skill level as it gets more information.
I’ve never coded a match maker, maybe I should, but Nano (another forum user) definitely gave it a shot.
Digital Signal Processing (DPS), Signal to Noise Ratio (SNR), and other software radio systems are pretty mathematically algorithm heavy.
You know what, I just had an idea. Given how scary accurate DeepMind was in predicting the opponent’s moves in chess, go and Sc2, I wonder if similar kind of A.I. can be used to improve the algorithmic matchmaking. A.I. could potentially foresee even more normally unexpected things like people leaving games, throwing, picking the hero that they never play etc. by gathering all kinds of information about the player’s habits. That and tightening the SR gap in matchmaking should result in more “fair” matches in ranked imo.
Jeff Kaplan stated in a dev update (years ago at this point) that AI is currently not implemented in Overwatch match making but they were hoping to incorporate it in the future.
But in short, you bet it would help
This is where business and tech use lingo but maybe have different visions or meaning.
If AI being machine learning
Machine learning being the process of having a value but having uncertainty and adjusting it.
Mmr is a form of machine learning / AI in OW1.
You can’t tell me Mmr is not a form of machine learning (also a core reason why I think it’s a classification system not a progression, machine learning is all about classification with certainty)
So what folks actually mean like bots?? In the game? Or a more one depth machine learning for player placements… who knows since he’s not even in the company anymore.