Seriously, matchmaking is horrible. 0 synergy and unbalanced teams.
And not necessary because you get burned by it, even when you win the games are a mess.
No communication, one side fest either way. No fun.
The game disgusts me atm. Took a break, came for 5 games, all in which I had 0 fun because the teams were completely unbalanced.
RiotBazerka14 votesCurious if you had any thoughts around how we could design a system to account for things like how often a player communicates etc. Sounds like a lot of the factors that make up "balance" as you describe it is more around the people playing the game that you have to play with rather than the game itself?
DeathBurst5 votesYeah. Hence the S to "gameS", as in, matches. And the explicit mention of the "matchmaking".
GripaAviara4 votesOk here are my 2 cents for the matchmaking:
get rid of MMR ELO system, because obviously getting rid of the **league ** system is not in discussion: I get it, ELO system is easier to balanced mathematically since there are tons of algorithms out there, but this "hack" never worked. The answer to this is sure: "but dude, this will leave too many gaps, like how do we account for boosters? how do we balance quick climbing, or "going down" . Here comes my next point
lane mastery matchmaking : You already have this system in place, it's right there in the game where few know how to find it. You give a score and account for a number of parameters for lane prowess. So you compute a 0-1 scale mastery that is ever-changing and visible. This will have 2 outcomes: people will try to improve their actual game, not just chasing kills and ofc in matchmaking.
How will people then not abuse this by having one max lane mastery and 0 in others and they will matchmake? If the difference is too big and he is not autofilled, the game accounts the lane mastery+(abuse rating 0.5-0.8 scale)*diff(max lane mastery- lane mastery) .
What about the player pool? +2 and -2 div will mean a player pool of 5 divisions. Master + Challenger should be locked together + if it passes a certain time to account lower division (like: 5 minutes passed, consider D1, 10 minuts passed consider D2, etc)
5050BS3 votesMake People Climb from Iron4 up. LP gains are flat not scaling. If you win you go up. We dont care about boosting Smurfs to higher level faster. YOU can tell who a smurf is.
If Im winning 50.5% of my games then Im climbing in all ELOs. If Im 49.5% then Im falling. Its EZ to see and figure out. Not this Im at 54% win rate but hard stuck because RIot shit system says I should be falling and gives me 9w/14l LP shit.
Kai Guy2 votesIf Im 49.5% then Im falling
Your currently at 47% of your games. So in your system you would be iron 4 right now?
Edit. I really have no idea why you want to be ranked at the bottom of Iron 4 and I do not remotely understand how or why you think your suggestion makes for any improvement to MM. So... business as usual for us both then?
5050BS1 votesIm at 47% because Riot keeps putting Iron players on my team.
Ill go 15/4 and they will be 2/12
This happens over and over because at some point in the Troll teams I had Riots "System" has said "Oh now you are bad and must fall"
Even tho Im -17/+19
Its like Riot is wanting me to carry these Iron players but its impossible to carry them.
Since Riot says I should be falling I am falling. Not because of my skill. My skill is much better then who Im grouped with most of the time. These teams are uncarryable.
Kai Guy1 votesYou don't magically stop getting those players. Population is still going to distribute around What the average player looks like. The most common level of skill is going to be the "Average" level of skill. So above and Below Average are lower population sizes.
Changing the Title makes no impact on population distribution. The starting value is where average will pool.
So downgrading Silver into Iron makes no real impact and would strictly be cosmetic. It also would make you a iron 4 player as your currently sitting in Bronze1-Silver4 MMR match averages, the current 50% marker for player percentiles is around Silver 2-Silver 1.
Does seeing this with a visual aid help you understand it?
Kai Guy1 votesHow much better would a +2 player be vs a 0. How much better then a -2? What odds would you give them?
GripaAviara3 votesI meant there to be the player pool and those are the division range they are matchmade, like a Silver 1 can only play with Gold 4 and 3 and silver 2 and 3 . Thus the matchmake pool is G4,G3,S1,S2,S3 .
Now matchmake pool is based on MMR, division don't really matter just as an approximation for the hidden MMR.
Ideally, an incrementation system would be the best, so the pool starts with your division only, then in 30 seconds widens one more +1 and -1 division and in 3 minutes widens more
Kai Guy2 votesOk, But the assumption is that higher = better right? I need to know if your trying to do it subjectively or objectively for your ratings. So can you put a expectation on how much better +2 is compared to a 0 for example and is it consistent between players?
What is the point to division +1 or -1 if they are not separating players by skill as You cant use that to have MM for skill if its non existent. So, how do you grade skill and what does it look like? Are you talking about an Objective system or Subjective system for grading?
If you use WR and you have set expectation for probabilities to win? That's an Elo derivative. If you want to use a Subjective rating that ignores Win rates and who your vs that's a Subjective system. Something in-between can be done, if that's what your aim is how much better are players across divisions? What value is that and how is it generated?
Please tell me you have some numbers in mind and solid logic to support it. I want to know what your talking about here. Elo? Subjective rating? Other (If so what is its values?)
Also how does this fix non communitive teams or deal with some one who tilted then threw a game?
GripaAviara1 votesif they are not separating players by skill
Being in a division is a separation of skill in itself. Then you have the lane mastery rating I proposed for a deeper matchmaking in that skill range. Now the division distribution is indeed skewed atm, but if you reset the ladder and implement the system, the skill will redistribute itself
The rest of the questions make no sense to me, so please rephrase them more clearly
Kai Guy1 votesYour not super clear on what the system looks like player end. Same as it currently is? So Iron to Challenge?
Then are your using a distribution for 5 divisions that's hidden behind that system or not? that's the -2 thru +2 range you mentioned.
Riots League system has 0 information on skill with out MMR. You want to change the distribution Model so I want to know what are you using for that. How much better does your system expect a -2 vs -1. -2 vs 0? Etc. MMR and Elo use their distribution model to build and set a value for this, when that value is wrong they adjust.
What is defined as more skill vs less skilled by your system? If your unable to create a consistent value for this, a reliable expectation of skill for how often you expect player A to beat player B. How can it be used for MM?
Subjective judgment is looking some one and going they did X Y and Z very well, that's 8/10 points! So that would be observing for context of each game by holding up a standard of how closely did some play to the correct move all game. The problem doing this is if you don't have the ability to watch and grade with the context of individual games subjective rating wont work well because 1500 damage on an adc in game could be a game winning result due to short game time compared to 2500 in game B which would be low due to say game time.
Now you can try grading on a curve. Take average stats and individual stats to check for positive or negative gap. But average stats in league are gonna to skew and population and popularity will constantly be shifting your expectations. You also have a very real issue, not all champions need the same metrics. So you cant run it by just Role averages, how does some one with 0 CC or 0 healing generate those values? I hope its self evident why you'd have to account for champions individually.
So if we get to the idea that winning = skill and who your playing = how hard it should be to win or lose? Thats Elo. Its not subjective. It does not care about gameplay, just the results of a game. Did you win or lose. Who are you playing. But you don't want to use that and you think it will not work? "Remove Elo/MMR." So my attitude is if that's what your asking for.... remove Elo/MMR then replace it with an Elo/MMR. Seems a bit pointless yea?
How does your system generate its values? You provided some undefined values are saying that hey if you use this it will work. Why will it work? How are your systems values generated. If you cant provide WHY they work is there any reason to not treat it like a random guess on your end where you just pulled out some numbers because they looked good? What does your system do that's superior to the work of mathematicians and statisticians for a well established and peer reviewed system?
Ways I know about and what uses them. Normal Distribution and WR? Elo. Verhulst population logistical distribution and WR? Elo. That was a 9/10 play? Subjective grading. Your X stats compared to the average Y and wins do not matter. Like Fantasy football I guess, Statistic grading. Your X stats compared to the average Y and direction of movement is decided by Wins or loss? Elo.
GripaAviara2 votesSo,
It's pretty clear there won't be a perfect system. But the current one is definitely not working. It hasn't worked for years, but when solo carry was more impactful and the games where less snowballing skill alone could save games which made teams less relevant than today. Now, people can see the flaws of the MMR system, especially in lower than diamond leagues, where the "population" is huge compared to the end of the "league"
Now you question is how does the system generate the values, the system already gets those values https://am-a.akamaihd.net/image?f=https://news-a.akamaihd.net/public/images/articles/2018/august/stats/Stats_Combat_EN.jpg
Are those parameters absolute and clear? No, definitely. But they are the best you can get without AI. A further study from big data Riot gathers can reveal other parameters in the future.
So why rely on them then? Because they are the league average, let's call them Skill Performance Indicators (SPI).
What is the +2 and -2 div ... it's the player pool. Why this pool:
Why keep the SPI in the first place? Because while the division is an average skill indicator the game is structured in 5 different roles with at least 2-3 playstyles for each role. So you need a deeper grading for better matchmaking.
Also,** very important**, the SPI must be based on pool average otherwise you can have skewed results favoring underperforming and disfavoring overperforming, something you don't really want. You want good players to climb, and bad players to demote. You also need a way to track him swapping to abuse the system, I called this abuse rating. You simply track how often he swapped with his lower positions and how big is the difference.
SPI should be calculated based on Challenger best median values over 20-50 games
How will the system works abstractly:
Let's take John, he is S1. First the pool is created: S3 S2 S1 G4 G3 From this pool, the games compute the average pool SPI for John's selected positions. The SPI scale is 0-1. Let's say John has a jungle SPI of 0.2 and an ADC SPI of 0.3. The pool average SPI is 0.4 and 0.5 for these positions. John occasionally has and abuse rating of 5% of his matches, so his abuse score is 0.05. His highest rating is the top with 0.4 SPI. Because John has an abuse score of 0.05 this means 0.05 * 2 (<- a constant that Riot should use to adjust the abuse prowess) = 0.1 (the score is capped at 1) of the difference between the highest is added to other SPI. So his new SPI are Jungle: 0.2+0.1*(0.4-0.2) = 0.22 and ADC: 0.3+0.1*(0.4-0.3) = 0.31
Now let's return to the system. A quick recap: Jungle SPI: 0.22 , ADC SPI: 0.31 and Average Jungle SPI: 0.4 and Average ADC SPI: 0.5
The game now creates a team. He looks for other people with John SPI for Jungle since is his primary role. But because of John's SPI being bellow the average for league, the game will look to matchmake him with someone who has a better SPI than John by a factor of a difference. The constant should be adjusted by Riot after field test. How will that work:
John Jng SPI: 0.22 - person to matchmake John should be at least 0.22+ 0.4 _ (<- average jng SPI) _ - 0.22 * Skill Gap Constant (let's take it 0.25) =0.22+ 0.18*0.25 = 0.22+ 0.045, so the game will search around 0.265 with a small variation (let's say +-0.005 or +-0.01, i dunno, these should be tested in field)
The game has a hard time to find a jng in those values, tries on ADC. Computes the ADC SPI and finds one.
Voila, you'll have the whole team nearly balanced.
For further balance, they can factor the fact someone is below or above the median SPI and try to balance the number so that each team has the same bellow and above players
Hope this clears the air. Imho this will be a way better system. Not saying there aren't other better ideas out there, but Riot should make their mind and implement one already
Kai Guy1 votesAight. So 0-1 scale is equal to 0-100. For folks who don't like decimal points. Honestly not remotely important. If some tells you your system cant work from your scale value. Its an arbitrary value so ignore any comments that call out your scale as a problem. What generates the values is whats important.
I dont think a hard reset will help as you have no "challenger" to build The curve and using prior seasons would just make it mimic prior seasons. Assuming you put directional weight on win or lose.
If you do a weight off w and l for direction your basicaly just asking for modfication to riots current mm system. As they do mmr by role and "abuse" ties the range for all roles tighter together.
If you want weight on more then W/L I do not know a model to use besides AI for automation. Your system is weak if it can't see what stats are impactfull or just noise.
Personally I am biased to put weight on wins loss cuz it cant be faked with ease. Stat padding is weak, hurts teams and fucking easy in a mobas. Holding skills for proper usage is correct playing Stunning a Kat r for example but raw stat as a the metrics reward mindless spam. I dont know how to make a non subjective system that can track this with out wins and losses as a core value.
I might be in error on things at times and I will correct when it's proven. I have so far attempted to be clear when a system is mmr or elo so you dont feel like I am trying to trick you into admiting elo and mmr work. I would rather show you the logic behind them. The math that builds them and how to test them. Help you fill out your own system then compare and test the two.
So for your curve. I can give you some models for building curves when I get home. Did you have a distribution in mind?
GripaAviara2 votesI am biased to put weight on wins loss cuz it cant be faked with ease
It's already weighted in win and lose, the system doesn't renounce that. You got the league system that creates the pool. You can increase the "weight" of the win-lose weight by decreasing the pool range from the start. Win-loss is not that relevant in current league as a matchmaking tool because the solo carry is limited unless a heavy smurf. So you could be great and loss a lot of games because of bad luck.Sure you will eventually climb but it will be a grind fest
Current stat system is pretty decent. Also you need wins . If spamming stats give you win then good job.
I have no distribution in mind
Kai Guy1 votesWell then. Sounds like you want a mmr system with stat adjustment. Ever look at the publications for trueskill?
Far as distribution goes. That's how you build team odds and move accounts at what speed you want for rate of change. Logistical has slightly different curves at the tail end but they wont be to dramatic in an Elo system. Fractions of a % for the majority.
I got access to a bunch of formulas. You may want binomial distribution at the start.
GripaAviara1 votesIt's not a true MMR system because you don't have a definitive player rating you matchmake around. Not, it's a matchmaking system so yeah they are similar i guess in scope.
TrueSkill2 is interesting and could be an improvement to the current algorithm, but alas, is patented and I dont' think it will work that well with the league system.
Kai Guy1 votesWant to point out that theres already issues in riots system from class intervals not being respected for their duoing system which creates visibility mismatched games. Its even more pronounced for Flex and Normals.
So.... Rounded everybody's MMR. Which represents a range. Into their title, which represents a Range of ranges. Makes for lower quality MM. Think about if you treated a high uncertainty Gold 1 player and low uncertainty Gold 4 player as exactly the same and equally fair to put on a team.
That's what I see it doing, Blending a range into a much wider range. I don't feel that's remotely the right steps to take for better MM Quality and it still wont fix the gameplay issues generated by player behavior. Curious why you feel it would be stronger. Looks to me like your making σ A weaker value, so your standard deviations get all screwed up.... which is what sets We and K. So you harm the systems ability to compare 2 low uncertainty accounts with accuracy. Player A is expected to win 4/5 games vs Player B. You can do this with an Elo system, generate exact odds. Then when the odds do not get met? The system adjusts.
How do you generate odds with no definitive Ratings? Set Value per Rank+Division? Youd go from being able to match teams and have an exact MMR average between them to having a unknown value for where in the range some one was.
Well, that's how I would do it, that's the issue I see with that method. You have a better implementation?
If you like TrueSkill 2 here is a relevant quote for SPI
For example, consider a team game where players cooperate to achieve a goal. An improperly-designed skill rating system could encourage players to impede their teammates. As another example, consider a system that increased skill rating according to the number of times a player healed themselves. This creates an incentive for a player to repeatedly injure themselves so that they could be healed. In general, the more control that a player has over a quantity, the less useful it is for skill rating.
A lot of folks are going say Nah dismiss TS 2 because it was generated using FPS games. But the principles can be adapted and optimized. As Long As Riot develops their own metrics specific to their game I doubt that there is much Microsoft can do.
Hordes664 votesWow, after 8 months a rioter finally shows up to acknowledge the existence of a thread about the awful state of season 9 matchmaking. Don't you have balance issues to ruin?
Kai Guy6 votesWow and players wonder why reds dont bother with gameplay boards.
GripaAviara1 votesWas surprised too, but a welcomed one
The thigh guy2 votesCurious if you had any thoughts around how we could design a system to account for things like how often a player communicates etc. Sounds like a lot of the factors that make up "balance" as you describe it is more around the people playing the game that you have to play with rather than the game itself?
If games weren’t so snowbally it would be much more fun. Most games are decided at 10 minutes, if that.
Comeback mechanics are a joke and turret plating just added to the snowball mechanics in the game.
Zullar2 votesWhat metrics does Riot use to determine how "snowbally" the game is? Can you share your data?
-How often do people surrender @ 15 -How often do people surrender @ 20 -How often a player goes wild with their K/D/A and gets like 15 kills with 2 deaths. -What the typical spread of razed turrets is (i.e. does one team kill 11 turrets while the other team only kills 0 or 1)
I bet if you looked at those metrics you'd see that league is currently very snowbally compared to where it was.
Additionally look at champ design. 5 years ago lets say you had a really fed 10 kill Morgana. Well... you could play safe and try and avoid her binding, maybe even buy a Banshees, and do your best to avoid getting bound by the really fed Morgana. If you got bound it was your fault and you paid the price.
S9 if you have a really fed Akali/Qiyana/Any new Riot champion that has 12 dashes, 3 shields, 4 invisibilities, 7 stuns, 2 heals... you cannot avoid them. If you are anywhere on the same screen as them they can engage on you and kill you regardless of what you do. There's no counterplay. This makes the game feel terrible.
The mobility and burst has really started to ruin league. Most new champs feel toxic with all their mobility and overloaded annoying kits. -Zoe -Yuumi -Qiyana -Akali rework
Look at Kha'Zix. He has 1 leap that resets on kill. If he jumps in and assassinates he can jump out. But if he fails to secure the kill he's stuck in the middle of the enemy team and dies. Seems like a fair design.
Now compare to Akali...
Beerstein1 votesCurious if you had any thoughts around how we could design a system to account for things like how often a player communicates etc. Sounds like a lot of the factors that make up "balance" as you describe it is more around the people playing the game that you have to play with rather than the game itself?
I think there is a bit of a fix but not in the sense most think. Matchmaker takes into account your account stats, not champion stats. This can cause huge discrepancies.
So the goal would be two fold.
Number 1 will be a product of 2. So you can't know what a player will pick. But if you instead apply the same logic matchmaking uses to match players AFTER the picks and bans you can generate an MMR modifier. Essentially it'd use account stats to match, and champion stats across both teams to rate the final impact on your MMR.
What this will do is tighten up the swing in either direction from mismatches resulting in slightly better MM. But it will also be less punishing on players who justly feel the match is hard to reach in regards to winning, and also more rewarding if they do.
The result in better mm would likely be small, but the loss or gains to MMR might make people feel their MMR is more in their hands despite massive imbalances that are indeed out of Riots control as you don't know what champion players will play.
Sorry for any typos, am on phone.
P.S. Thanks for participating in the forum, it's much appreciated. A lot of us have a lot of thoughts that we feel are never heard. Even if it's not your job, we appreciate the dialogue. I've worked in gaming a long time and a little communication goes a long way.
Karn Bishop1 votesNobody is really asking for a system like that they would just like to not be in a match where everyone on their team is high silver/mid gold except for the adc who is bronze 4 or the opposite problem of having a ridiculously high ranked player vs people the system should never match him against.
The drop in match quality after MMR was hidden has become more noticable as time goes on to the point where even the laymen are noticing the decline in match making quality.
It was always a fear of mine that MMR was hidden for the main purpose of being able to play fast and dirty with match making and sadly that seems to be the case especially with the announcement of eternals which encourages the behavior which was used as a reason to remove being able to see MMR and for data sites to see said data as well.
