Elo Formula

[CaptainNeutron]

With that said, you could use the old Elo values as starting point, they will be more accurate than resetting every vet to 15...

[jgbaxter]

QUOTE (tmc @ Sep 2 2006, 08:18 AM) But why? Mathematically, this makes absolutely no sense. This simply makes the total number of ELO points go down, continuously. The player's ELO cannot mathematically converge (and I assure you, this is a mathematical certainty) unless the average player's ELO stays constant (and in this case, 1500). If the ELO always drop, the ranks cannot and will not plateau.

I see your point of course, unfortunately I can't see a way that will make that better then what's currently in use... I would be alright with the losers losing a proportional amount instead of the absolute elo.

As it is, like previous versions, elo will (for the majority) continually drop for players... the only thing holding them up really, is that their age rank occasionally increases... then even that changed yesterday...

It's all very frustrating. /smile.gif" style="vertical-align:middle" emoid=":)" border="0" alt="smile.gif" />

[Flower]

Regarding the converging to a certain skill plateau:

As TMC said there is the syndrome of a elo rating deflation with symmetrical ELO implementations, due to the fact that old players that leave tend to take with them more ELO points than the new players that join add to the current "ELO Point Pool"
To combat this there are commonly adjustments to the ELO gained or lost...
I think the dev team implemented them with the *4 to *0.5 modifiers that JG wondered about further up (look at the elo code at the top :-)

So it might be that the devteam already succeeded to prevent the deflation to a certain degree. Only time will tell of course.

Though even if Skill deflation were to occur, Elo is in its design a relative skill rating system. If you would run a Elo calculator on a database containing only Grandmaster chess players (and assuming that base skill is 1500) they would be rated significantly lower than the same calculation applied to a database containing masters and intermediate chess-club players.

Elo just tries to predict the winchances by relative comparison /smile.gif" style="vertical-align:middle" emoid=":)" border="0" alt="smile.gif" />
Of course the Elo can be used as an absolute skill measurement (as your position in relation to all other objects around you can make an absolute estimate). But this absolute measurement is just a byproduct of the relative ranking system and thus it is not necessary for skill to 'plateau' in order for elo to function.

Well of course it would be better if it 'plateaued' But its very difficult to adjust the system so that it does.

P.S.: This Elo deflation problem is known to cause trouble with players that are inactive over long periods as their rating would be relative to the 'old playerbase' instead of the 'current one'.
Some chess clubs try to offset this by adding a temporary elo penalty (until they play more frequently) in order to counter this problem and prevent the "I am sitting on my high elo scores and thus am king of the hill"
symptom

[Raindog]

QUOTE (CaptainNeutron @ Sep 2 2006, 06:59 AM) With that said, you could use the old Elo values as starting point, they will be more accurate than resetting every vet to 15...

I think the age rank should be considered at the time of a reset. This lets you get more accurate information quickly. As time progresses and ELO become more accurate, simply reduce the effect that age rank had until it is just ELO. Basically if a 30 age rank adds 1500 elo to start(30 * 100/2), after 1 week it adds 1000(30 * 100/3), 2 weeks add 750(30 *100/4), etc. Age rank is more accurate than ELO rank during the first week and becomes less accurate as more games are played so just design the system to take this into account.

[jgbaxter]

Hmm, dropping it after a reset as time progress makes sense, still, I'm a big fan of having the age rank avergae with a players elo rank, it mitigates the extremes and gives a somewhat more accurate gauge. /smile.gif" style="vertical-align:middle" emoid=":)" border="0" alt="smile.gif" />

[tmc]

To combat this there are commonly adjustments to the ELO gained or lost...
I think the dev team implemented them with the *4 to *0.5 modifiers

That is incorrect. As I understand it, these modifiers work in both directions (up and down), and hence have no real bearing on the total amount of ELO per player. While it does decrease the number for the vets, the effect itself is pretty much invisible. What this really does is count top vets less than they should. when it sees someone who should be, say, at 3k points without this adjustment, it will consider him to be at, say, 2.8k, and does so because it doesnt agree that the player is worth 3k points (ie, 2 1500-pointers).

[Flower]

QUOTE (tmc @ Sep 2 2006, 10:00 PM) That is incorrect. As I understand it, these modifiers work in both directions (up and down), and hence have no real bearing on the total amount of ELO per player. While it does decrease the number for the vets, the effect itself is pretty much invisible. What this really does is count top vets less than they should. when it sees someone who should be, say, at 3k points without this adjustment, it will consider him to be at, say, 2.8k, and does so because it doesnt agree that the player is worth 3k points (ie, 2 1500-pointers).

Hmmm ok, if these values have not been introduced into the equations in order to combat elo pool deflation then it might be worth to add them.
(any of the dev team who could confirm my or tmc's (and JG's) interpretation?

How to scale these values in order to stabilize the pool is beyond my basic knowledge of Elo though.
Lets hope that AllegElo will be refined even further if it proves to be inaccurate with the passing of time.

Till later *takes a look at the sql code again to see how the 4 to 0.5 modifiers are used*

[tmc]

It's impossible to scale these modifiers to make ELO/player constant. All you can do is make sure that the same amount of ELO is gained than is lost, and that new/leaving players dont unbalance the equations.

[Fragtzack]

Stats based rankings is extremely better then Elo. Elo system is one on one relative comparison. Individual Elo has no place in a team based game. Elo in Allegiance "penalizes" players regardless of the players skill or how good the player played. What Elo for Allegiance does do is encourages stacking, because the only way to gain Elo is to choose the team with the better pilots/commanders.

Some of the best players in sports can be on losing teams all the time. However, because of thier stats they are recognized as the best. In Allegiance Elo system, its all about picking your team to gain rank. Allegiance Elo is not truely reflective of pilot skill, or anything close.

Elo for Allegiance has huge problems. The number one problem with Elo for Allegiance is the encouragement of stacking. The other problems are all over these boards about how the system is not working and will not work.

Stats based rankings had only 1 issue: scout whores are not awarded for naniting/probing.

The stats based system does award for winning games. You get double points for being on the winning team. So both the winners and losers still gain points (ranks). The winners get proportionally more points. All that is needed is a way to award points for naniting and probing.

Let's spend energy fixing the stat based system that has 1 small problem instead of trying to implement a square wheel called Elo. Elo in Allegiance is simply a measure of how good a pilot is in picking the team to join. Elo in Allegiance has very little to do with individual pilot skill.

[Dengaroth]

QUOTE (Fragtzack @ Sep 3 2006, 05:30 PM) Elo in Allegiance "penalizes" players regardless of the players skill or how good the player played. What Elo for Allegiance does do is encourages stacking, because the only way to gain Elo is to choose the team with the better pilots/commanders.

...

Some of the best players in sports can be on losing teams all the time. However, because of thier stats they are recognized as the best. In Allegiance Elo system, its all about picking your team to gain rank.

This is the main issue right here. Your ELO rank isn't something you should be aiming to "improve", like you do with stats. It's trying to evaluate your impact. By trying to "gain ELO" you're intentionally trying to overrate yourself. Of course the system won't take too kindly to that.

QUOTE (Fragtzack @ Sep 3 2006, 05:30 PM) Elo in Allegiance is simply a measure of how good a pilot is in picking the team to join. Elo in Allegiance has very little to do with individual pilot skill.

If only even games count, and you've been on the winning team for 800 of 1000 of those even games, isn't there a certain statistical probability (converging on 100% as the number of games increases) that you are contributing more to your team's winning that someone who's been on the winning team for 200 of 1000 such games?

Notice the "even game" prerequisite. Right now, after a reset, this criterion is a little skewed because the system reset to a state where it's assuming all players are equal to one another, and will need time to stabilize again (which is the main reason why I think resetting the ratings was stupid in the first place).


I do agree with you on one point, though. There should be a traditional, point-based system implemented, simply because some people feel a need to "climb rankings"... which a rating system won't accomodate. Such a ranking system won't be usable for balancing teams at all, though (imagine someone who only turrets. All the time. Constantly. Now compare to a bomber pilot. Or a scout pilot. Or even someone who kills miners in scouts all game).