still really unbalanced

[RageRitual]

Got games like this all day today:

We still really need to balance the 10*s

43* vs 48*

Genx 7.5
lockdown 10
saiyan 10
sunny d 7
zapata 8.5

VS

best 10
booker 8
iron survivor 10
mythril 10
racka 10

NEXT MATCH 36.5* vs 43*

Cape 8
GenX 7.5
Klean-X 7
midoent 7.5
tj hazuki 6.5

VS

Best 10
BOOKER007 8
Exalt 7.5
Mythril 10
rylo 7.5

[Rab]

This is mmr, twdt stars do not apply. Can't rly complain if the mmr is balanced. More games required, or initial adjustment to mmr to maintain known ranking.

[Riverside]

Those lines are what twdd and twjd lines have looked like for months.

At least with ?go twdd3 and twjd3, there's a medium-term and long-term opportunity at balancing teams from a systemic POV.

[Exalt]

I don't really see a problem here? MMR doesn't care what *s are. It balances based on MMR score, which goes up and down based on how well people play together. Obviously, it still needs more time so some players drop and others rise to where they belong, but that's something everyone knew about beforehand.

Also, here's another game to counter-balance your point: https://twd.trenchwars.org/showgame/90125076

Name	W	L	R	R1	R2	R3	R4	KO	TP	LO	RES
Best	5	10	64	2	2	1	0	Lockdown	6m 34s	0	1888x1016
BikEy_BoY	6	10	75	2	3	0	1	iron survivor	10m 6s	0	1920x1080
Exalt	8	10	80	5	1	1	1	iron survivor	9m 4s	0	1920x1080
Mythril	8	10	94	1	6	0	1	Lockdown	11m 25s	0	1920x1088
saiyan	9	10	83	2	7	0	0	iron survivor	8m 29s	0	1920x1200
Name	W	L	R	R1	R2	R3	R4	KO	TP	LO	RES
iron survivor	11	6	139	2	6	3	0	-	11m 25s	0	1920x1080
Lockdown	15	8	143	4	9	2	0	-	11m 25s	0	1920x1200
Siaxis	12	5	121	3	7	2	0	-	11m 25s	0	1360x1024
Suav	4	7	88	0	1	2	1	-	11m 25s	0	1920x1200
Sunny D	8	10	95	1	4	2	1	Mythril	9m 18s	0	1920x1080

That's about.... 47* or so vs 42* or so. 42* lineup won. Again, *s don't mean everything in MMR. Sometimes it just has to do with ability to team together or one or two top players having a bad game, which happens to everyone. MMR is more of an "average" thing anyway, so off games and great games average together to create the actual score in the end. Off games like this are expected in a statistical standpoint.

Also, you complain, but what you experienced happens to everyone. i got a string of the two worst MMR-rated players in the game in my lineup for many games when it first started against two of the highest rated, over and over again one night. Everyone started at 3000, so this has to happen to get people to where they belong. Play more and you will work through it. Larger sample sizes is what's needed.

[RageRitual]

I know these matches arn't based on DT ratings, but I am using those ratings as an example to show how unbalanced some of these games are.

The problem is mostly the 10* players, which are the best in the zone in their ships lose a game early on and are seriously underrated creating really unbalanced lines. A possible solution would to start people at different ratings based on their DT ratings.

Another example of how bad this is, roiwerk has at the time of this rating 1977 rating in jav. Which led to yet another really unbalanced line.

I get what the bot is trying to do balance the MMR scores, but this alone is not working. This is leading to hilariously unbalanced lines over and over. We need more variables for the bot to balance lines with.


Riverside The point of MMR is to balance lines, comparing this to TWD makes no sense.

Exalt I am not complaining for the sake of just being a bitch, waa waa I lost some games. I am complaining in hopes to open dialogue up for possible solutions. You say you have similar problems that just proves my point. At the start when I was winning every game I was still complaining it was unbalanced.

One possible solution would be to start players ratings based off of their DT ratings. Every other game is as bad as this:

MatchBot12> [MMR] Rating: 3232 -> 3147 Change: -85 Games played: 21
[TWDD4] Freq 1 vs. Freq 2 30 - 50 30 Sep 2022 21:31:56
,-------------------,-------K----D---R1-R2-R3-R4---Rtg--LO---Note------------------Time--.
| Freq 1 / 20 | 38 | 5 11 2 2 | 7 | 1 | Lose (-20) 7:21 |
|-----------------' | | | | | |
| GenX 7 | 8 | 0 4 1 2 | 7 | 0 | |
| rEnZi 4 | 10 | 1 3 0 0 | 0 | 1 | KO > dreamwin 7:02 |
| Sulla 4 | 10 | 1 2 1 0 | 0 | 0 | KO > Mongrel 4:38 |
| TPZ 5 | 10 | 3 2 0 0 | 0 | 0 | KO > Kassius 4:34 |
+-------------------,---------+----+-------------+-----+---+-----------------------------+
| Freq 2 / 36 | 20 | 13 20 2 1 | 59 | 0 | Win (+20) 7:21 |
|-----------------' | | | | | |
| dreamwin 10 | 7 | 4 6 0 0 | 15 | 0 | |
| iron survivor <MVP> 10 | 4 | 4 5 1 0 | 17 | 0 | |
| Kassius 7 | 3 | 3 2 1 1 | 13 | 0 | |
| Mongrel 9 | 6 | 2 7 0 0 | 14 | 0 | |
`--------------------------------------------------------------I=IN---O=OUT---#=DEATHS---'

2 people out 4 minutes, even Ravage lasts 5 minutes what the actual fuck?

[Exalt]

2 people out 4 minutes, even Ravage lasts 5 minutes what the actual fuck?

Like I said though, the more games everyone plays, the less bad teams happen. Since everyone starts at 3000, people naturally have to rise or fall over time based on their abilities before matches are truly going to be even all the time. I will say that there are a lot more even matches in MMR that I've seen even at the early stages than bad ones, but even the bad ones serve a purpose. It's only 1 round, and while it may suck to be a part of those bad games, they help eventually lead to better ones.

Also, doing things based on TWDT ratings would remove the entire purpose of MMR to begin with, as MMR is supposed to take out the human biases and errors and make things completely based on ability to win alone. Stats don't matter, biases don't matter, * caps don't matter, etc. Sometimes you have to learn to win with people you consider bad. That might mean changing your style up to actually win a game instead of just getting the best record or staying alive, etc. I'm not saying that game you last posted was winnable, but strictly on paper, it wasn't an automatic loss if everyone played well together.

Anyway, even if ratings were adjusted for TWDT and artificially fucked up the point of MMR, hypothetically they wouldn't matter past a game or two anyway, as people win/lose 80+ points each game. A couple rounds could skew the results for everyone negating hours of pointless work. MMR only works over a long period of time, not a tiny sample size. It's one of the reasons that make it good in the first place. Even an entire TWDT season is only what, 10 rounds or so total?. That's not enough games to accurate rate anyone statistically with MMR. 10 straight seasons of TWDT is still only like 100 rounds, which is more where MMR wants before it will be accurate. 1000 rounds is even better.

[qan]

The MMR system was built with the idea that we'd play qualifying matches. Thus the large changes in the first few matches. We realized that was just going to be too tough to try and do, though, always needing to group matches by either quals or regular matches. Since we're running unrated players with those who have played many games, the amount of change for early ratings needs to be reduced, because it's not having the desired effect of quickly doing a "best guess" sorting unrated players.

(Side note: It could be possible to take into account the uncertainty of ratings of players on both teams. If ratings are less certain, ratings changes probably shouldn't be as large because they're not as indicative of actual skill. Some games already do this. The math would need to be done properly, though.)

The biggest problem, though, is that we don't have the population to run quals, unlike every other game. It's a unique problem and may need a unique solution. Using DT stars as essentially a replacement for quals is one. Mostly we'd be talking about fairly trivial starting ratings adjustments. For example, starting people at 2500-3500 based on stars would be a meaningful change that wouldn't upset things too much. Using past data to start players at a rating in a new season is used by many games, especially as a way to sort players for quals. In our case, yes, stars are human data and are prone to error, but if star influence is fairly small and only used as a very basic starting point, it could allow for better balance over fewer matches played. In the end, ratings will even out to a distribution of a player's real estimated skill over time, as mentioned above, so a slight starting adjustment doesn't really tarnish results. (Also, remember that unlike Elo, MMR is not zero sum.) I'd understand if people are opposed to that on principle, though.

[Exalt]

I'd understand if people are opposed to that on principle, though.

If enough people want it, I suppose it's not an awful idea, but I don't see how that could happen at this point without resetting a lot of people's played games, unless you only do it for those who haven't played any thus far. I guess we aren't that deep into it yet, but I believe I'm around 40 something for wb and 20-30 for jav. Much of the community is around 20 or so played games. I guess if the decision is to ultimately change things based on * ratings, the sooner you reset scores and do that, the better, as I'm sure multiple players might get upset if they end up playing 100-something games only to get reset because a few complainers who played only 5 and didn't like the results.

[qan]

It'd be possible just to pre-load those who have under X games (probably 5, in keeping with the idea that it would effectively replace quals) and then set them at 5 GP.

The problem with needing to fine-tune the ratings adjustments for early matches does mean we'll need some kind of reset at some point. However, players who've played a certain number of games probably wouldn't need a reset, as it could be assumed that the rating was starting to approach their real skill range. That may be the compromise. In the end, this is still a testing stage, so there may be a need for some kind of reset at some point (whether partial or full), but I'd like to avoid it. We'll need to see what things are like once we're running automated games and all the major kinks are worked out.

[2pac]

qan killed the bot