Does Riot use deep learning algorithms for their matchmaking?

i cant move wtf·Dec 20, 2017·3 votes·1,266 views

I feel like this would be a great application for the technology. Maybe a rioter can shed some light on this? I think it would be an interesting discussion.

11 Comments

bywords70Dec 20, 20172 votes

i don't see the application, at ALL

mathematically it is simple, there is no need for deep learning algorithms considering we already know the answer to the problem.

the problem occured ever since riot put team builder into ranked and now we have a trade off of lower queue times and balanced matches.

the fact that you are even suggesting deep learning for a problem we already know the answer to just shows how you don't understand how deep learning works at all.

i cant move wtfDec 20, 20171 votes

LOL if you don't see the application then I suspect that's a deficiency in your own imagination, not my understanding of deep learning. I think there's a good chance they already do employ it for this purpose to some extent but it would be cool to see some of Riot's perspective on it regardless. Thanks for your contribution!

bywords70Dec 20, 20173 votes

you mean you probably failed 5th grade math and don't understand that matchmaking is a SOLVED problem that we mathematically know the solution to.

you're basically asking for deep learning on the quadratic equations.

you truely don't understand what deep learning is all about, it is basically an extension of human intuition. for example, if you were to model something like stock prices, we already have models of those where we humans put in the weights to variables that we think is right but deep learning is able to refine those models and put more accurate weights on those variables. in deep learning, all we give are the rules and the computer figures out the model with no human input. essentially the machine is learning like a human for a problem, but it has way more computing power. if our brains were more powerful then it would get the same results as deep learning.

in stockfish vs alphazero, stockfish has variables in it assigned to by humans while stockfish assigns its own variables based on what it thinks its best in self-play, same thing in alphazero vs alphago, in alphago, its moves are already contaminated with biased human moves while alphazero has no human input in it and it puts its own weights on evaluations of positions.

i cant move wtfDec 20, 20171 votes

Again you see the failure is your own imagination. To do matchmaking purely based on mmr, I agree it's a solved problem. But a deep learning algorithm can take into account much more "fuzzy" variables that likely do affect how the game plays out. Things like what types of champions the player plays and what role they got. Perhaps even win/loss streaks. These extra variables are difficult to quantify but no doubt have a bearing. A deep learning algorithm that takes these into account could potentially create games of even matching well beyond what a pure mmr algorithm could produce and could likely match teams with wider mmr gaps while still maintaining competitive integrity, which would lead to faster queues. I think it's interesting and worth exploring and I would be kind of surprised if Riot never looked into this or doesn't use some of it already.

bywords70Dec 20, 20172 votes

I guess you don't understand that your type of thinking is already the stockfish way of thinking. You are making up variables that YOU think affects the outcome of the game. You clearly don't understand that deep learning is NOT for solved math problems and NOT for variables that humans make up.

MMR is already the solved solution to the problem, it takes into account all variables that affects the skill level of a player if our definition of skill is based solely on winning and losing. Therefore any deep learning algorithm will do the exact same thing as MMR. Matchmaking based on purely MMR is the unbaised estimator, anything else is not.

You only need ONE variable to determine the outcome of games, winning and losing. everything else does not matter because mathematically speaking the summation of the weights on those variables are already being taken into account using MMR. We do not know what those weights are but we do NOT care. MMR is ALREADY the solution that deep learning will come up with because the problem is so simple its SOLVED.

i cant move wtfDec 20, 20171 votes

In a pure 1v1 game where you always get your "role" I would agree. But there's no feasible way in this game to just come up with the weights for things like "putting a player at such and such an off-role when their main role is X" or be able to match wider mmr gaps (having a player on your team with a significantly higher mmr probably matters more if they're say... ADC vs supp - not sure if that specific example is true but I think you see my point).

And I'm not saying the current algorithm with human defined weights isn't fair, it is. Would an adaptive deep learning algorithm be less fair? Maybe... but Riot also doesn't release their algorithm. Or would it be able to estimate a person's true mmr much more quickly and it would take less games for a person to "be at their elo"? I can see Riot not wanting to do it for that reason alone.

Again I'm not bashing a "flat" unchanging mmr system. I just think the questions of matchmaking integrity and efficiency of placing a player at their elo in less games using an adaptive matchmaker that goes well into the nuances of what actually happens in a match (players not getting their role, whether some types of players are more compatible with others for who knows what reasons, how far you can actually stretch an mmr gap before the disparity between role power really starts to become significant, etc) are interesting to think about. Is it even ethical to have an adaptive matchmaker in a competitive game, even though it's still technically just an algorithm? I think there's good discussion to be had here.

bywords70Dec 21, 20172 votes

You are already missing the point of matchmaking. Matchmaking is NOT designed to make games have a 50/50 winning chance. You could put 2 challengers on each team and rest bronze and that would have a 50/50 chance of winning. Matchmaking is designed to accurately estimate your skill level in the least amount of games possible, that is a solved problem. In fact when the teams are made you can calculate winning probabilities based on % pick rate of champions that people play and their % win rate against the enemy champions, that's just simple statistics.

If you want each team to have an actual 50/50 chance of winning that would require more computing power but still DOES NOT require deep learning because that can be solved by known statistical methods.

If you want a super computer to analyze one of your games and estimate your skill level based on each action that you make in each frame, yes you can use deep learning for that but that is pointless because we have MMR which requires way less calculations. The purpose of deep learning is to use LESS computing power for a result, which is why the point of alphazero vs stock fish is to show that alphazero looked at WAY less number of positions. For go its to show that alphazero can logically prune positions and reduce the number of positions it analyzes only the ones that make sense.

You cannot justify matchmaking having giving you a 50/50 chance of winning a game. Example, let's say currently you play a champion that gives you a 50% chance of winning against 50% of the champions the enemy players, and you lose 100% of the time against the other 50% of champions played by enemy players. This means matchmaking MUST match you against champions you have a 50% chance of winning against if that's what you want matchmaking to do. Overall your mmr would stay the same despite you obviously being a shitter winning only 25% of the time against people of your mmr.

What matchmaking will do is force you to lose until you play worse players and then it would be like a 50% chance to win 75% of the time, and a 50% chance to win 25% of the time. This is the best estimate of your skill level. All those other factors you are talking about like people not getting their roles are ALREADY factored into mmr.

Also I should say riot does NOT want to accurately measure your skill level in the least amount of the games possible, otherwise we wouldn't have this shitty league system.

i cant move wtfDec 21, 20172 votes

You're forgetting that a pure mmr system is only informed by wins/losses. Again I don't argue that there's anything inherently unfair about this. The point of any matchmaking system is absolutely to match you against an opponent, that based on your match history, you will have a 50% chance of winning or losing against. But there's actually quite a lot of information beyond simply win/loss history that can be used to inform matchmaking. Just look at all the data in the year in review Riot recently put together. There's much more than that even.

It would be an insane pursuit for a human to attempt to give all these other variables accurate weights in the context of matchmaking. But it's not unreasonable to assume a deep learning algorithm could... the data set is huge. Whether that brings up ethical concerns about competitive integrity or would bring new and old players to their correct mmr in less games and would therefore be divergent to Riot's own interests I think are the more interesting discussions.

Mysticman89Dec 20, 20171 votes

Keep in mind things like individual champion skill isn't something that can affect matchmaking, since matchmaking occurs before champion select.

Theres certainly ways they could look deeper into things and predict the outcome of a match once it's started, and once all runes are picked and champs locked in such system might predict one side heavily favoured to win, but that sort of thing isn't useful to matchmaking.