How exactly do we expect low rated players to play the game?
-
@clyf like I said above since you only have data on the performance of the team not the individuals, since there is no 100% correct algorithmic way to tell how much someone contributed to the win. So to rate a single player accurately you need them to play with different teams.
So to find the correct rating for a player in team games, that player will need to play more games than they would need to if they only played 1v1. -
You need to play more team games to rate a player accurately, but playing more team games doesn't guarantee you can rate a player accurately.
More games is not the issue, I think everybody recognizes that. How the system models how an individual player contributes to the success of the team is the issue.
If who your teammates are is the dominant factor for success, "your" rating will converge on the (shifting) average of the performance of who you are playing with.
-
@clyf said in How exactly do we expect low rated players to play the game?:
If who your teammates are is the dominant factor for success, "your" rating will converge on the (shifting) average of the performance of who you are playing with.
But who the deciding factor is differs between games and as there are bad players in both teams the team with the less bad one is still more likely to win (assuming the "good" players are about equal).
The trueskill test was also done on halo 2 data and a player having 0 kills and 1 death probably less of a contribution than a player with 10 kills and 0 deaths, but their algorithm completly disregards that and only compares team performance and from that infers the players individual new skill estimate without creating some kind of player performance estimation from the game data. And according to their tests that worked and the ratings converged.Though in FAF, if the rating discrepancy becomes too high and it would be straight up better for the lower players to give their base and not play, then it becomes hard at evaluating their contribution, but as long as they are contributing their rating will over time adjust to their real rating. Depending on teamsizes and the actual contribution they had this might take non feasible amounts of games though. But that's basically like a Halo player just sitting at the spawn to not die, which would also prevent the system from accurately rating him. (Unless you consider that this was just the optimal play in that situation so he is actually not that bad)
The problem is that it is simply impossible to design a rating system that accurately evaluates a players contribution without affecting how people play the game in unintended ways, unless your game is completely solved and you can precisely rate each move at each step. Unless you know the perfect path to victory you can't tell by how much a player diverged from it during the course of the game.
-
the team with the less bad one is still more likely to win
This is the part I don't agree with. At a certain point bad players drop off the bottom of the influence spectrum.
-
@nex said in How exactly do we expect low rated players to play the game?:
Though in FAF, if the rating discrepancy becomes too high and it would be straight up better for the lower players to give their base and not play, then it becomes hard at evaluating their contribution, but as long as they are contributing their rating will over time adjust to their real rating.
@clyf said in How exactly do we expect low rated players to play the game?:
This is the part I don't agree with. At a certain point bad players drop off the bottom of the influence spectrum.
yes, but it's like that in any other game too
-
No, because in the example of team slayer, bad players will continue to contribute (negatively, most likely) to the outcome of their team no matter how bad they are.
-
@clyf they could just stay at their spawn and do nothing, that might even be better than bad players hogging eco in FAF.
While not every bad player has "influence" on the game win (I think even in FAF bad players have at least some influence on the game even if they don't feel like that) they do have the opportunity cost of having someone better in that spot.// Even read in Discord recently from some ~2k player (was it zwaffel??) that being matched together with a 1k against 2 ~1600 is just unwinnable. This is 2v2, but it translates to teamgames, just their contribution will be much lower.
-
I apologise for the following misunderstanding if people are talking about trueskill 2 here... I tried searching!
TrueSkill wasn't just invented to judge the teamslayer game mode. What made people think that? Are you sure you're not talking about trueskill2?
I've looked into trueskill (but admittedly not trueskill2), and the internet search results I'm putting in now about it are fraught with explanations that really seem to miss how much the ELO system is already doing!
There are straightforward ways to get a normal distribution out of the statistics, but without doing anything needlessly complicated (and ultimately pointless) you can summarise what trueskill added to ELO as simply as:Trueskill was designed basically to be the ELO system, with 3 major differences (and they really do 'oversell' how different it is from ELO, imo!)
- One "change" (if you can even call it that) was to have the 'k factor', (K-factor: the uncertainty used in ELO systems, essentially how much you're 'gambling' on the current game, typically higher for new players, lower for veterans, to help zero-in on skill quickly), to be a lot more granular and dynamic (many chess organisations did this already, to varying extent, so it's hard to even call this a trueskill change, but trueskill does seem to introduce an enlargement of the K-factor from unexpected results, so I'm giving them credit there...). This really isn't as clever as it sounds - you just go from a k-factor of '24 for a new player, 12 for someone we know', to 'K-factor is 24 to start, goes down by 5% every time a match ends, unless it didn't end how we predicted, in which case it goes up by 5%'... That level of simplicity works just fine (though microsoft 'tarted it up' a bit.)!
- Another, rather simple change was to factor in team games (And it's really not hard to make the ELO system do this - we're talking addition and division with an adjustment factor based on rating 'spread'; though many games also add a separate factor for premades.)
- Third (and I think this is both the biggest, and most 'brilliant' change): it was designed to make players 'feel' better about their rating, by ensuring that it trends upwards at the start of a player's career - and this change is even easier to achieve - you just hide their actual ELO, and instead display whatever rating you're relatively sure they're better than! (We've already done all the work with the dynamic K-factor, which as I said many ELO systems also already did!)
It isn't some special complex system that microsoft designed purely for teamslayer Halo games. Now it wouldn't surprise me if a certain gamemode used a variant, but it would shock me if a game used anything other than victory, defeat, and draw as possuble after-game result (pipe up, I'm eager to learn!)
Trueskill is essentially the ELO system with a bit of stuff that tons of people were already doing with the ELO system, and an added ego boost!
@clyf Trueskill2 is doing more, factoring in more than wins and losses, from my understanding (I haven't seen the maths behind it). Again, I apologise if this is what was being discussed and I missed it! Could that be what you're talking about here?.
As for why supreme commander can't base skill rewards on things other than games won; @nex summed it up well when talking about what behaviour you want to reward players for, aka what the 'intended' behaviour you want to encourage is.
It's true that basing the system on a factor less binary than 'win or lose' would probably zero-in on a rating faster than this, but it would be (quickly) 'zero-ing in' on the WRONG rating.
Rewarding the idiot that gladly suicides their entire army and commander to kill the red-health ACU that was about get strat-bombed regardless?
(Or so many similar situations) : you don't have to play many moba games to realise just how eager players are to act this way if a number next to their name is involved!I've written rating systems for team-based strategy games before, and tried all sorts of metrics. Games won has to be the only relevant metric for a few reasons, but the best I can do if those explanations don't feel like 'enough' is just saying I promise you - I've tried other metrics, they don't represent actual skill nearly as well.
tldr: There's no point 'zero-ing in' more quickly if we're zero-ing in on the wrong rating.
(And tbh, part of the success of the trueskill system is down to how players have more fun during the zeroing-in process! So it doesn't really need reducing in many ways!)
-
@sylph_ The teamslayer mode was just an example where different rating mechanisms (KDA) might be applied. I also only specificall referred to the tests they did with trueskill on the Halo 2 dataset (including team games) that were mentioned in the paper.
No idea what they currently use for team slayer or what trueskill 2 is.
-
Even read in Discord recently from some ~2k player (was it zwaffel??) that being matched together with a 1k against 2 ~1600 is just unwinnable.
If it's impossible for a ~3000k team to win against a ~3200k then the score isn't a good predictor of team composite performance. Do you understand that?
-
@clyf that is not the fault of the rating system per se, but more of the different environments the players play in. because if they'd play more games together then the ratings would start to match their real skill for that gamemode again. It's the same how some players get to 3.6k rating.
Rating only really makes sense if there is sufficient exchange between the whole playerbase.
And of course the high volatility in the performance of a low game count 1k player. might work well might not and also 50% winrate is considered an autoloss for the average player since for a game to feel balance you need a 66% winrate, which is impossible for everyone.
But in this game the low rated player has a high impact on the game, if it's actually unwinnable is something that can and should be questioned. -
@nex said in How exactly do we expect low rated players to play the game?:
No idea what they currently use for team slayer or what trueskill 2 is
Trueskill 2 is a proposed system that includes more metrics (like kills) than just win/loss/draw in its statistics considered.
(unlike trueskill)
which paper you read might be very important here, since trueskill (1) has been released, and assigns skill based ONLY on whether a game was won, lost, or drawn (no KDA or ingame score or frags or anything).
Trueskill 2, I believe, still has its maths 'hidden' (and all the overhyping that such a move enables!) This might have changed since I last read about it a few years back, though.As I understand it, trueskill 2 is geared towards FPS games with scores, which is why I thought it might be what was being referenced here regarding 'teamslayer modes' and the like. At the very least, it's confirmed to take other factors besides a binary victory/defeat into account.
-
that is not the fault of the rating system per se, but more of the different environments the players play in
If there's one environment where the player is rated at 1000 and winning 50% of their games (steady state, 4v4 with 4*1k/4*1k rated players in game), and another where a player is rated at 1000 and getting crushed (2k,1k/2*1.5k), and no further distinction is made between those two environments, then their score is converging on a different point for each environment, and no amount of games played will result in their score converging on a single point.
(emphasis for emphasis, not to be a dick)
-
@clyf my point was not that the statement "2k/1k vs 1.5k/1.5k is unwinnable" is true, but that the opposite of your statement "low ranked players have no influence on the match" also exists and the truth lies somewhere in between.
The reason most 2k+ players don't play with <1.5k players together is that it immediately ruins the game, which would not be the case if they had no impact. Lower rated players are just a lot more inconsistent than high level players, which makes a game more of a coin flip. -
My original statement in regards to how TrueSkill evaluates expected team performance:
At higher level games, where players have the knowledge and wherewithal to identify and exploit weaknesses, it seems the weak link is the deciding factor.
At lower level games, where players are on average less able to identify and exploit weaknesses, it seems the strong link is.
To refine the entire point: TrueSkill assumes that the relationship between individual performance and composite term performance is mathematically linear, while empirical evidence in FAF suggests that it is not.
-
Okay, here's our problem--
I said:
your kills and deaths being reflected directly in the game deciding score
You said:
I can tell you that is wrong. The only relevant metric is you won
What you thought I was saying was:
your kills and deaths are used to calculate your TrueSkill rating
... whereas what I actually meant is:
Your kills and deaths [are summed with the kills and deaths of your teammates] to determine the score [which determines who wins the game, which is then used to calculate TrueSkill]
At no point did I say that any metric other than win/lose was used in the TrueSkill calculation.
-
@clyf said in How exactly do we expect low rated players to play the game?:
At no point did I say that any metric other than win/lose was used in the TrueSkill calculation.
Yes, but you suggest that this is a metric that shows how good you are / how much you contributed to the win, which is false as there are parts to the game that are not reflected in that score. (like killing an enemy about to kill your teammate, which in turn generates more kills)
@clyf said in How exactly do we expect low rated players to play the game?:
To refine the entire point: TrueSkill assumes that the relationship between individual performance and composite term performance is mathematically linear, while empirical evidence in FAF suggests that it is not.
As soon as teamplay is involved the performance of the team is no longer simply the sum of the individuals, otherwise it wouldn't ever make a difference if you play with a friend vs. you playing with random people on your team.
But assuming a linear relationship is fine as it evens out with enough games, unless someone intentionally plays with higher rated players and takes the spot with the least influence on the game, but that's a problem of custom games in general. Sometimes your spot makes the difference sometimes it doesn't. Sometimes you are even the high rated player in a lobby.The rating system makes a statement about the statistical distribution of wins/draws/losses over time. So a game can feel absolutely unwinabble if you play 1k/2k vs 1.5/1.5, but that is just part of the randomness involved and you get that feeling because suddenly your own contribution is lower than you expect since a lot hinges on the fact if the worse player has a good day or not.
-
metric that shows how good you are / how much you contributed to the win, which is false as there are parts to the game that are not reflected in that score
You don't understand what a metric is.
Are you Armistice840 on the discord?
-
How the hell did this thread veer so far off course? Shoo.
-
I wanted to downvote you but I'm going to make the right choice and bail on this whole conversation instead.
