Maybe a hiden separated queue when the deviation of elo of premade teammate are too high.
Yea a lot of games use similar methods to keep premades stomping solo queuers.
But i dont think its something that we would see in this game. Those systems are complex and need constant fine tuning.
Just use the correct elo addition method. Then naturally the players with the highest elo have the biggest impact on the team elo and it will be very hard to smurf-push them.
Like i said. New system does not affect badly solo queuers, nor does it affect premades with similar personal ELOs. New system only affects badly those who make premades with ridicilous ELO caps.
Thatās just wrong. it will affect all players, as this system will naturally push all players towards the average team elo and therefore underrate the top players and overrate the low elo players.
Teams will be extremely unbalanced and the expected winrates of the team elos will be extremely inaccurate. Teams will be matched that are heavily disbalanced towards one side but assume they have (50:50 winrate). And top players will probably only team up in only high elo teams as they would know if they let one low elo player in they will lose a lot of elo on average. Leading to top elo teams that will likely get 70 % winrate or whatever and also need that winrate to hold their elo.
Remember: The elo system was invented to match players of equal strength. Thatās the basic concept behind it.
To my surprise i read that the devs finally tried to fix the rating system. But not to my surprise they messed up again. This thread contains multiple solutions, but the devs thought it was better to pick their own one.
The elo calculation should be a zero sum game, but the devs made a system that isnt zero sum. This really looksl ike an idea of the devs without any real thoughts. I think we still should use this thread, since the devs still didnt really adressed this issue.
devs made a system that isnt zero sum
Not saying they are right, but how is this system not zero sum? As long as both sides gain/lose an equal total amount of elo.
Iām not sure, but I think there is a way to exploit the new ranked elo system with clever matchmaking.
There are certain team matchups were you can get net elo. If you manage to make this kind of team and match up against your friends and intentionally get a certain win/loss rate you should be able to push your teammembers.
It needs probably some kind of rotation in your team composition, but when itās figured out I see the chines guys doing it that way againā¦ without the need of any smurfsā¦
One example:
Game 1:
Player A | Player B/Y | Player C | Player D | Sum | ||||
---|---|---|---|---|---|---|---|---|
Elo | 2400 | 1800 | 2000 | 2000 | ||||
Real Elo | 2400 | 1800 | 2000 | 2000 | ||||
Shown Strength | 13.54160626 | 7.059153677 | 8.77118932 | 8.77118932 | ||||
Real Strength | 13.54160626 | 7.059153677 | 8.77118932 | 8.77118932 | ||||
Current Team Elo | 2100 | 2000 | ||||||
New Team Elo (as it should be) | 2786.43003 | 2638.412146 | ||||||
Winrate vs other Team | 1 | 1 | 0 | 0 | ||||
Winrate exp o Player | 0.909090909 | 0.240253073 | 0.359935 | 0.359935 | ||||
Gain per Win | 2.909090909 | 24.31190165 | 20.48207999 | 20.48207999 | ||||
Lose per Defeat | 29.09090909 | 7.688098347 | 11.51792001 | 11.51792001 | ||||
Net average elo gain | 2.909090909 | 24.31190165 | -11.51792001 | -11.51792001 | 4.185152549 |
Game 2:
Player Y | Player B | Player C | Player D | Sum | ||||
---|---|---|---|---|---|---|---|---|
Elo | 1800 | 1800 | 2000 | 2000 | ||||
Real Elo | 1800 | 1800 | 2000 | 2000 | ||||
Shown Strength | 7.059153677 | 7.059153677 | 8.77118932 | 8.77118932 | ||||
Real Strength | 7.059153677 | 7.059153677 | 8.77118932 | 8.77118932 | ||||
Current Team Elo | 1800 | 2000 | ||||||
New Team Elo (as it should be) | 2438.412146 | 2638.412146 | ||||||
Winrate vs other Team | 0 | 0 | 1 | 1 | ||||
Winrate exp o Player | 0.240253073 | 0.240253073 | 0.759746927 | 0.759746927 | ||||
Gain per Win | 24.31190165 | 24.31190165 | 7.688098347 | 7.688098347 | ||||
Lose per Defeat | 7.688098347 | 7.688098347 | 24.31190165 | 24.31190165 | ||||
Net average elo gain | -7.688098347 | -7.688098347 | 7.688098347 | 7.688098347 | 7.10543E-15 |
If a team consisting of 5 players (A, B, C, D and Y) repeatedly plays about 40 % of Game 1 and 60 % of Game 2 (Game 1 with exchanging player B and Y every 2nd game) the team will gain about 1.67 elo per game. Most of it (1.16) will be given to the best Player A.
Elo grinding we have found you
(Best ratio is about 38.5 % of game 1 and 61.5 % of game 2 in this example, but itās probably not even the best example.)
In Game 1 why is the sum of net elo gain not 0?
Because of the new calculation method.
It is complicated to explain exactly the details and why it is positive in this case (itās also negative in other cases). But itās just what happens - and it is independent of which team wins.
Just if some team constellations are paired with other team constellations there will always be a net gain or loss (and it is constant for the team pairing).
And the way I constellated it I made it abusable to grind elo for a whole team.
The issue mainly arises if there is a big skill gap between players at the same team. If all are about the same, the sum is almost a zero sum, so it doesnt have much of an impact.
Just consider the following:
Team A:
1x 2000
3x 500
Team B:
4x 875
Both teams are on average 875 rated. If team A wins this game, then they gain 86 elo. Team B will loses 64 elo. So just this match will inflate the elo by 16.
In this example both teams are equally rated. Things can even be more off if the average of both teams also different a bit. It seems to become even more bad if the team with a high elo difference has the highest elo and wins the game.
And that is what happens a lot by premades. They sometimes have big skill gaps, leading to more issues on the current ladder.
The maths is correct
But I find very difficult that 1 player with 2000 Elo lose against 4 players 875 Elo
Agree with you that the Elo system has to avoid inflation in any way
Been playing around with different team game formulas, was planning to interrogate the data later to get some more evidence based metric but Iām curious as to what people thing fair āteam elosā should be.
For example say we had a team consisting of 600, 400, 200 Elo Players. What would you say a fair team Elo should be? If we were to say its the mean (i.e. 400) then that would imply a fair opposition team is 400, 400, 400.
I just tried an alternative Elo formula of:
b = 20
scaling = 400
x_scaled = x / scaling
log(mean(b^x_scaled), base = b) * scaling
Which gives an āfairā team score of 485 which I thought seemed reasonable to me, but yer I wasnāt sure what peoples gut feelings were on what would be a fair opposition team mid point ?
check this out please:
Like everyone has said before me, the total amount of rating gained by one team should be opposite of the total amount of rating lost by the other team. Else, rating will quickly inflate or deflate, like what we are seeing now. Furthermore, I agree that everyone in a team should win/lose the same number of points. I think @Haladon demonstrated very well what is wrong with the current logic of the rating system. Theoretically, this could all be solved in the following way: Both teams should be aā¦
O for sure, sorry the point of my post wasnāt the specific formula I used above but was to try to probe people on what a āfair teamā looks like. In the post you linked he left the specific scaling parameter blank with the caveat that we need to find a good value for it which I was trying to inquire out of people here. I could convert my above formula to exact same one used by him but the question still stands
EDIT: My plan is to use the match data to see if I can get an empirical value but was curious to see what peoples gut instincts are i.e. for a team of 1200, 1000, 800 what do you think a fair opposition team who are all of the same Elo would be ? 1000? 1050? 1100? 1200?
EDIT: Also do people think this scales linearly i.e. if the fair value for 1200, 1000, 800 is 1100 does that mean the fair value for 2200, 2000, 1800 is 2100 ?
If you want to figure that out you probably need to analyze game result data either with some kind of multi-dimensional correlation parameters or via hypothesis and chi-square test (smoothing probably required).
I mean it would be really cool, but as we just had a reset it probably again needs some time to stabilizeā¦
1200, 1000, 800 is 1100 does that mean the fair value for 2200, 2000, 1800 is 2100 ?
Well that is what the elo system is designed to be. (Itās not linear btw, itās logarithmic)
If people donāt think it should be like that, they have a wrong perception of elo.
But of you could make a poll to figure out what people expect and then try to chose the lambda value accordingly. But if peoples perception is too far off you will possibly get a value too much out of the ācorrect oneā. If itās a vlaue cloe enough in the proximity cause of the counterforce of the individual elo calculation it will result in a quite stable distribution, but if itās too far off depending on matchmaking and so on peoples elo probably will fluctuate a lot - and result in incorrect win predictions again.
If you want to figure that out you probably need to analyze game result data either with some kind of multi-dimensional correlation parameters or via hypothesis and chi-square test (smoothing probably required).
I mean it would be really cool, but as we just had a reset it probably again needs some time to stabilizeā¦
If the current calculation is off be a lot, then the ranking will never stabilize at all. As result you cant really draw conclusions.
I dont know if the devs have still the outcome of all games? If so, then they should recalculation the elo of all players for all matches based on different elo calculations. And then compare the outcome of these different methods based on some statistical tests. Not sure how viable this idea is.
I was thinking of something along the lines of a logistic regression model that uses something like either the difference in Elo dispersion within a team or the difference in max Elo from the team average as a predictor variable (along side standard difference in mean Elo and civ selection variables)
Assuming this comes out as strongly predictive of the result I am thinking you can then apply different team Elo formulas with different constants until you can fit a model where the Elo dispersion is no longer a predictive factor which would indicate balence.
Iām then thinking you could repeat this process for different slices of the data say 800-1200 and 1400-2200 to see if the same formula provides balence across all Elos.
Though Iām not super confident in this approach so if anyone had any ideas to refine it I would appreciate it