Star Rating is even more broken now.

I just joined a game immediately before it ended, giving me just enough tome to join the losing team and have the initial class selection screen appear. I didn't have an opportunity to play.

I lost .8 rating.

What the hell?


Star Rating doesn't matter that much, but really now. This didn't happen before the last update, and if it's ever going to be useful at all for matchmaking/team balance...

EDIT: Played another short game. I played maybe two minutes of what was likely a 5-minute game at most, based on the MVP stats. I went 5 and 2 with 20 bot kills and lost another .3 off my rating. It looks like rating no longer pays attention to time at all, making it literally completely worthless. :|

Try playing at least a couple of full games before complaining for a change. I know it's super hard but try.

That's the thing, though - when you play a full game, rating changes are better IMO than they were before. It seems as though things are only worse now when you play an incomplete game. This specific problem has never happened to me before, and I've definitely joined numerous games that ended pretty much immediately throughout my playtime.

Also... joining a game and not being able to do anything at all should not cost a player 0.8 points from their rating. I don't see how it's unreasonable to ask for a change when something seems to be that far off. I know a lot of work was put into the last patch, and I appreciate what's been done, but something still seems wrong.

I'm also fairly sarcastic and cynical.

Let us know how it is after at least 10 games.

I have played at least 30 games since the update :lol:

Overall the star system is much better. But I get the feeling that changes to your rating are based on the MVP in the server. Every time I am MVP I go up anywhere from 0.1 to 0.4, any time I am not the MVP I go up either 0.04-0.05 or I go down.

One thing that should not happen: I had a 2.5:1 kdr with a decent number of bot kills at the end of the match, but a guy on our team (MVP) went like 18-0 so my star rating went down.

I don't know if it really is based off of other players in the server, I could just be paranoid here, but if this is the case it should really be based on the individuals' performances.

Solid update, loving Funland and Chickey. Keep up the good work!

Star rating is fine.

You didn't lose 0.8 star rating 'cause you didn't have a chance to do anything. You lost 0.8 star rating 'cause that game knocked off your oldest game that was being accounted towards your star rating. And that game was one of your better.

. . . that suggests that one's star rating only accounts for the last x games, which doesn't seem like a good way to rate performance to begin with - it's more work to record and average performance over the last x games (and only those games) than it is to alter a number after every match based on a formula.

Ratings should NEVER be based on individual performance unless there are objective tests to qualify you (ie, skill tests)

It is to rate you quantitatively relatively to everyone else.

Back to OP:
I've lost 1.0 star from joining a game as the team lost.
0-0-0-0-$350

gg

I've played well over 10 games since the update, and until now, nothing seemed to be wrong. The changes to your rating after a full game seemed to fit just fine, and seem to pay less attention to KDR. I've been playing with clanmates, so all of the games had been played through in their entirety up until these two earlier, which stood way out.

That could explain why my rating dove so hardcore. The MVP went 20-0 in that first game I mentioned.

Time needs to be taken into account as well, then. Losing 0.8 rating after not being able to do anything while most of the people there (including the MVP) played a full game is quite skewed.

There are also some things that are objectively important to the game. Kills, deaths, assists, bot kills, turret kills, and damage to the ball are some examples. The only way to base your rating change on everyone else and their performance is to also weight your rating against theirs, which means the rating needs to judge their skill adequately first.

Right, in that case rating would be based on relative performance based on set metrics, probably kills, deaths, assists, bot kills, money. What I was saying, is your rating should never be tied to a fixed metric number, ie, 50 pro kills should never always equate to +1 star, especially when you're pub stomping and the next highest rated player is 2 stars lower.

Apparently the former poster thinks he deserves a rating increase for a set post-game score no matter how inferior it is to another player or MVP (assuming everyone was in for the full game)

20-5-0-100-$4000 is great, it means crap when MVP is 50-30-0-200-$7000

Go into game. 2.73
Money Ball already damaged. bots swarming. half my team still in the spawn. I go out and try to stem the tide. Get two assists, one death. By the time i respawn, money ball is done.

My new score because of that game is 2.39.

Given that both players have the same rating going into and spend the same time playing in that match.

If A can frame 4 paintings in an hour, and B can frame 3, it suggests that one is better at it than the other and can therefore do more in the same amount of time. This is a PvP setting, though, so you have to base the value of kills, deaths, and possibly assists on the rating of the players you kill/who killed you.

EDIT: Also, to suggest that everyone should be relative to the MVP means that good players can potentially ruin everyone else's rating (unless changes take rating difference into account).

I completely agree... my point was that if it is based on the MVP in the server, that is not at all rating you relatively to everyone else. It is rating you relatively to everyone in that particular server. With a game like MNC where people skill levels are all over the place that just doesn't work as intended. On any given day the same person can join a server and lose 2 stars over the course of a gaming session, but the next server he joins he can be playing against supports that are trying to heal him to death.

It's just not a reliable measure of the user's skill (relative to the entire community).

My friends are convinced that the rating system is in relation to the MVP of the game, and they flat out told me that they do not enjoy playing on my team, because I am MVP +90% of the time

(That was my quote :< )

Thinking about it, the idea of managing to rate players' skill appropriately seems to be a very complex task. It needs to weigh their performance against the people they play against, but also consider general averages - if everyone does poorly, the best of the worst (so to speak) shouldn't gain much, if anything. For example, it seems like the average number of bot kills is about 4-6 per minute. If the average number of bot kills in a 15-minute game was 60, and a player ends with 80, they should gain a bit more than most people in that game. (More than 15*5=75 and more than the game average.) However, if the average was only 30, and someone ends with 40, they should see very little change if any from that part of their performance. (Less than 15*4=60, but more than the game average of 30.)

When you look at direct PvP stats, you also need to consider the ratings of the players involved. Someone who went 20-0 could have killed only players with ratings 2 stars below theirs. At the same time, someone who only went 5-10 while the MVP was 25-2 isn't doing too badly if all their kills were against people with ratings 1.5 stars higher and the MVP is rated 2 stars higher. To make this even more complex, if the MVP played a full game and the other player didn't, the number of kills they "should" have had needs to take that into account.

Short version: Ratings change needs to be a very complex system to adequately rate someone's overall performance and reduce huge jumps, IMO - especially if matchmaking/team balance is to rely on it.

There should be no problem with MVP being the rating anchor.

I think you two don't like the idea of a zero-sum ratings system? (Because the s<P is always true, if the average rating is undervalued, this trend will continue so long as the sample set never changes.)

It creates rating pockets, akin to clubs and national associations for activities, where a 1800 rating at your club translates to a 2100 rating elsewhere, and 1900 to the nation-average.

I tried to integrate this is my rating system proposal, but the problem you get here is aggregate rating inflation. Also, it's very easy to gain/lose more stars in this system. Especially if you kill-steal often, or often die against multiple opponents (ie, grapple kill-steal, or dying to a nub Tank's DB while a good gunner was the one who pinned you down.)

You'd end up creating far more volatile ratings.

Your rating is going to be inherently flawed consistently in a random manner based on the opponents you face. No one is perfectly rated. Thus, no one will ever get a proper rating adjustment facing these imperfectly rated players.

This system more or less needs a true anchor, the best player of ALL MNC players, whom all ratings are derived from.

MVP-based rating system tries to achieve this in a sample size, but is wholly ineffective unless the 12 players continue to play only each other. Really, any rating system you can derive will continue to hold this flaw, unless MNC derives your rating from a longitudinal study and rate you based on your trend of KDA/BK/cash per game and cross references it with every player simultaneously (thus your rating will be affected by players not even in the same game as you.)

The crux of the problem is that the rating currently only utilizes information from a relatively small sample size (12 player or less games), but is meant to rate you against the entirety of the population.

wut.

The amount of traffic a database of lifetime stats for all players would need to handle for that purpose is more trouble than it's worth.

Base the rating changes off of someone's performance versus the other players in the game and their current ratings as well as predefined, snapshot averages. It's far simpler and can gauge someone's skill adequately enough for the purpose of team balancing.

No, you don't. You generate snapshot averages based on what the community has proven to be repeatedly possible, then use those in addition to weighted averages from each game (the weight of which is based on the ratings of the contributing players and/or time spent in the game) to decide how each statistic changes a player's rating.

What those snapshot averages do is set a bar that has nothing to do with the players involved. The best of a group of horrible players doesn't get a good rating unless they can repeatedly do what good players of the past used to be able to. IE we know you can go 25-0 repeatedly, so if suddenly nobody can get over a 15-5, they don't get 5 stars just because everyone else is worse. (tempted to toss in some bold or underline formatting just for effect here.)

Try simulating it, you do get inflation unless you force weights to net out changes which increase rating volatility. You'd have to impose a k-value to both have zero inflation, and less drastic rating swings per game.

There's no k-value currently, or if there is, it's 1 star or greater, which makes the purpose of having a k-value pointless.

Also, in regards to latecomers, it's difficult to imagine a good system for it that can't be abused by players trying to "preserve" their rating.

So what do you do, weight the rating change they should have gotten by the percentage of time they were in the game?

Uber hasn't even fixed players dropping from a game and rejoining, and the server having no knowledge of their former skills/cash and its corollary, the $600 glitch. I doubt they'd starting tracking time spent in game.

I guess the easiest stopgap is to not give a ratings change for people joining within 5 minutes of the game ending, or even better, prevent players from joining a server 5-10 minutes after the game starts. People generally don't want to get tossed into a half-started match, it affects rating accuracy adversely, and causes a probably imbalanced game to end more quickly creating more lobbies.

That's exactly what I said you do. More specifically, you generate a set of formulas that roughly define what statistics you should be able to generate per minute of gameplay based on your rating. You use those to generate an initial change in rating, then you adjust that change based on the other factors I mentioned. As long as the base formulas do not change, it's simply not realistically possible to continuously increase in rating without increasing your statistics (kills/deaths/etc). In fact, the -only- way for anyone to gain rating while performing under their required level as defined by those formulas would be if their opponents all grossly outranked them, and the adjustments overcame their deficient performance - the only stats that could do this are kills and deaths, since these stats should be weighted based on the level of the opponents involved. Bot kills would be weighted based on the average number of bots killed by the group, as well as the time spent by the player in that match.

EDIT: This is a game that can be rated objectively to some degree. Very few people end a 15-minute game with more than 40 kills, and I highly doubt anyone can do it repeatedly. You also can't earn fewer than zero kills. Assists could be worth anything from 1/4 to 1/2 of a kill to mitigate "killstealing". You can't die fewer than zero times, and I have yet to see anyone die more than 30. In short, there are what seem to be clear ranges that can be divided into smaller ones to base performance on. At the very least, the minimums seem to be extremely obvious and easy to define. 0 kills/assists, around 40 deaths or so, 0 bot kills, and the minimum amount of cash that can be earned in x minutes of gameplay. These define a rating of 0. What maximums are defined and generated values to represent those that aren't already will define a rating of 5.99 (or just 5, perhaps).

You didn't quote the rest of that paragraph ;p

So right, I agree weights would work to net out any inflation problems. But you need to add a k-value, which does not yet exist for MNC.

Yes, with weight it's all fine, in fact that's the way I proposed it in my thread. Only difference is that I did not take into account player1 and player2's star rating when P1 kills P2 or vice versa, as that requires on the fly calculation DURING the match, which I did not suspect was practical on Uber's part. Alternatively, you could keep track of every kill each player had at the end of the match against each other, but I'm not sure if it that can be kept track of in the current iteration of the game.

Instead of a hard k-value, I had the rating change modified by the spread of star ratings (twice the standard deviation), so a soft k.

As far as objective measures of rating, it's strictly more than a regular rating.

In the end, the ratings are here not to measure how good you are (since that's an infinitely more complex task than the actual intent), but to make decently fair teams.

Lyrae, I think we mostly agree on the math, and we're quibbling over technicalities; I'd appreciate a look over my rating system proposal if you don't mind.