Ranked teamgame matchmaking and elo gain/loss could use some tweaking

[noBrainur]

Ranked teamgame matchmaking needs some tweaking. There’s probably 1 in every 5 ranked teamgames that are actually unwinnable, despite leaving/rejoining queue after a few minutes to try to ensure this kind of game doesn’t take place.

The elo gain/loss is also in need of tweaking. In the above game I lose 26 elo despite the match being objectively unwinnable. In the below game I win 3 elo despite the match being FAR more difficult than the above match was for lucassadid. Yet 30 elo was granted to lucassadid (who did actually play quite well, but still).

I think the ranked matchmaking should be thresholded by the minimum L2 distance between the teams. If the minimum L2 distance between the teams is too high, then the game should not even be launched unless everyone in the game was in the queue for 5+ minutes. By “minimum” I mean fixing an ordering of team 2, say (v1, v2, v3, v4), and then finding the ordering of team 1, say (u1, u2, u3, u4), such that (u1-v1)^2 + (u2-v2)^2 + (u3-v3)^2 + (u4-v4)^2 is as small as possible. This is a mathematical way of trying to balance teams by pairing strong players against strong players, etcetera. This is just a suggestion for starters.

[Ashinferno]

I’m so glad some one else in this forum cares about team ranked games. BTW I know that kronk guy. did he play good? once he suicided all his units under my TC and called all his teammates noobs and left.

[noBrainur]

I do care about getting some TG tweaks, and I’d prefer this discussion be focused on the matchmaking and elo gain/loss algorithms.

[Cajocu01]

@noBrainur

The elo gain/loss is also in need of tweaking. In the above game I lose 26 elo despite the match being objectively unwinnable. In the below game I win 3 elo despite the match being FAR more difficult than the above match was for lucassadid. Yet 30 elo was granted to lucassadid (who did actually play quite well, but still).

Yea its def a wall to solo Q in Tgs and get this types of matches often. And you loose the equivlaent of like 4 normal matches in elo points.

@noBrainur

I think the ranked matchmaking should be thresholded by the minimum L2 distance between the teams. If the minimum L2 distance between the teams is too high, then the game should not even be launched unless everyone in the game was in the queue for 5+ minutes.

I think the issue is it currently takes the avergae elo and even if it believes its “balanced” is not the case. 2 1600 and 2 800 Would average 1200s But most likely a team of 1200s will do better since prob the 800s players arent gonna impact the game much and its gonna end up being a 4 vs 2 while they get going.

[noBrainur]

@Cajocu01

I think the issue is it currently takes the avergae elo and even if it believes its “balanced” is not the case. 2 1600 and 2 800 Would average 1200s But most likely a team of 1200s will do better since prob the 800s players arent gonna impact the game much and its gonna end up being a 4 vs 2 while they get going.

Yes true, for those teams the difference between their arithmetic means is 0, which suggests that the teams are perfectly balanced. But their minimum L2 distance would be

sqrt( (1600-1200)^2 + (1600-1200)^2 + (800-1200)^2 + (800-1200)^2 )

= sqrt( 400^2 + 400^2 + (-400)^2 + (-400)^2 )

= 800

Now consider if the two teams were (1500, 1500, 1000, 1000) versus (1500, 1500, 800, 800). The difference between their arithmetic means is 100. But the difference between the minimum L2 distance is

sqrt( (1500-1500)^2 + (1500-1500)^2 + (1000-800)^2 + (1000-800)^2 )

= sqrt(0^2 + 0^2 + 200^2 + 200^2)

= 283

So the arithmetic mean says the first game is more balanced, whereas the minimum L2 distance says the second game is more balanced.

Note that neither will take into account that the difference between 1600 and 1300 may have more of an impact than the difference between 1000 and 700. But quantifying that would probably be time consuming and would produce diminishing returns compared to the simple change from arithmetic mean to minimum L2.

Note that the algorithm for finding the minimum L2 distance would (I believe) just be matching the two highest ranked players to form the first squared difference (u1-v1)^2, then matching the next two highest ranked players to form the second squared difference (u2-v2)^2, etc. I’m 99% sure this will produce the smallest possible L2 distance between the two teams’ elo vectors, and I would welcome a correction if I’m wrong!

[Elrys]

I’ve been advocating a minored-loss system for months now. I Can 100% relate to that experience, especially when you start in low elo.

Losses are too punitive. I can loose against 1200 and win against satfat the next game, all of this because of matchmaking. I’m not getting rewarded correctly for my level in team games, and this is what makes the TG experience so frustrating nowadays, if you want to be serious about it.
Because while I complain about bad teamates, I am effectively not questionning myself and getting better. If I knew these kind of games didn’t affect my ranking at all, I’d try stuff out or work on specific abilities. Right now I have to play at 200% capacity, to ensure a loss anyway.

This is not sane for the game, long-term.

[Melinoë]

Yeah I do hope something gets changed about ELO gain/loss at least, I think the current system makes it impossible to take solo TG seriously, playing with randoms already can be a shitshow as it is, but losing 20+ ELO for unwinnable games on top of it is insanely demotivating

[Urban_Hunter]

You’re suggesting a formula, thinking it will work. From my experience, I can say that the difference in ELO between players should be limited. It shouldn’t be more than 300. If the highest ELO on a teamgame is 1600, the other players shouldn’t have less than 1300 ELO. This is the result we should strive for.

[noBrainur]

That’s a good new suggestion to consider. It would definitely lead to more balanced games than the current method does. But it would also remove some games that would have been fun and balanced. In my experience games can be fun and balanced even with massive elo differences, such as (1500, 900) vs (1400, 1100). The important factor is that the players on opposite teams can be matched up somewhat equally, and that the differences in those matchups somewhat balance out.

Some notes to the devs who may be reading this:

The formula I presented earlier can be improved because it doesn’t take into account which team each squared deviation favors. A better formula is one that does not involve the square root, and uses the signum function (or equivalent logic) to give negative values to the squared deviations that favor team 2. That way the method can distinguish between the following two situations that the previous formula would consider equal, but which are clearly not:

Situation 1:
(1500, 1200) vs (1400, 1300)

Skill Diff = |sgn(1500-1400) x (1500-1400)^2 + sgn(1200-1300) x (1200-1300)^2|
= |(1) x (100)^2 + (-1) x (-100)^2|

= |10000 - 10000| = 0

Situation 2:
(1500, 1300) vs (1400, 1200)

Skill Diff = |sgn(1500-1400) x (1500-1400)^2 + sgn(1300-1200) x (1300-1200)^2|

= |(1) x (100)^2 + (1) x (100)^2|

= |10000 + 10000| = 20000

Note that the previous method would consider both situations to have a skill diff of sqrt(20000).

A note regarding the calculations of elo gained/lost:

When we are calculating elo gained/lost, all the players on each team probably should be gaining/losing the same amount of elo. Maybe we don’t need any ‘fancy-schmancy’ skill diff calculation to fix this. Maybe it’s just as simple as basing the elo gained/lost off of team average instead of individual elo. Whatever the fix, it is not right for a player to lose 20 elo for losing an unwinnable game, yet gain only 5 elo for winning a nearly unwinnable game.

[Urban_Hunter]

@noBrainur

When we are calculating elo gained/lost, all the players on each team probably should be gaining/losing the same amount of elo. Maybe we don’t need any ‘fancy-schmancy’ skill diff calculation to fix this. Maybe it’s just as simple as basing the elo gained/lost off of team average instead of individual elo. Whatever the fix, it is not right for a player to lose 20 elo for losing an unwinnable game, yet gain only 5 elo for winning a nearly unwinnable game.

I completely agree with your position, it is terrible that the player with the highest ELO in the match is on a team with players with the lowest ELO in the match and gets nothing for winning.

[Riba]

Think problem is mainly incentivising players to play ranked. Most people just don’t want to play ranked where map pool is just bunch of AoE2 Arabias. Add bans so we can play Nomad, Arena, Jotunheim, etc. How do you get good matchups with less than 18 people in queue for 75% of the day?

In my eyes Elysium, Savannah, Watering Hole, Marsh, Ghost Lake, Tundra and Oasis are just the same map. I liked AoE2 ranked where you often got to play maps and be creative on maps like Arena, Black Forest or Nomad. I am playing with randoms anyway, why cant I just use some funky strats on funky maps.

[Cajocu01]

@Riba

Think problem is mainly incentivising players to play ranked.

I agree a lot with this part.

@Riba

Add bans so we can play Nomad, Arena, Jotunheim, etc

This one not so much. (Edit: yes to abdn no to the niche maps part) Some maps simply are only fun when you dont know how to abuse them. For example, until last patch playing norse on nomad was handicapping yourself basically with how units spawned.

Some maps heavily favor some civs over others, like acropolis being amazing for those that have good siege.

@Riba

In my eyes Elysium, Savannah, Watering Hole, Marsh, Ghost Lake, Tundra and Oasis are just the same map.

I get where you are coming from and some of those maps play similarish, but you can divide the ones that are very open or closed.