webdiplomacy.net

First of all, Elo is a name - not an acronym so you it doesn't need to be all caps.

Like others have said, GR is quite simple to understand. At the start of the game, everybody 'bids' a fraction of their rating and that makes up the pot. So, the pot is bigger if you play against better people but your bid is always the same regardless.

@Octavius, regarding new players:

The system mitigates these issues in two ways:

1) Games with more provisional players are weighted less. Right now, that weighting is 2*p/7 where p is the number of ppl in the game who have played more than 7 games.

2) player ratings are more volatile for new players. The weighting for a game is inversely proportional to the number of games you've played. This allows players to rise (or drop) to their 'true' rating quicker.

RoganJosh wrote: ↑

Sun Jan 21, 2018 11:52 am

Octavious wrote: ↑

Sun Jan 21, 2018 11:13 am

The other player gets a significant boost to their webDip ranking under both systems, but the difference is how it impacts you.

Both of these are zero-sum systems, so he can't gain more than what you loose.

Not to bog things down but they aren't exactly zero sum. Expected scores are based on an 'exponential mean' rather than the arithmetic mean so the rating changes don't always add to 0.

A_Tin_Can wrote: ↑

Sun Jan 21, 2018 11:22 am

in Ghost Rating, losses against expert opponents and beginners are not rated equally.

Yes, they are. This is because GR doesn't have a constant k-factor - the ratings of all players in this game is an input into the k-factor, which cancels out the earlier weightings. Have a play with some numbers in GR, you'll find it behaves very differently to typical Elo.

Really? Thanks for letting me know.

Yonni wrote: ↑

Sun Jan 21, 2018 12:46 pm

RoganJosh wrote: ↑

Sun Jan 21, 2018 11:52 am

Both of these are zero-sum systems, so he can't gain more than what you loose.

Not to bog things down but they aren't exactly zero sum. Expected scores are based on an 'exponential mean' rather than the arithmetic mean so the rating changes don't always add to 0.

My bad, I missed that part.

So, the system has a built in inflation due to the exponential function I guess?

@RJ

Yeah, but I suppose the average rating could shrink as well in each game. At any rate, I imagine the rating inflation is dwarfed by the inflation caused by low-rated players leaving the site. We could curve everybody's rating to keep the site wide average at 1000 but I'm not sure if that's necessary.

@Yonni

Since the exponential function is convex the exponential mean is greater than the arithmetic mean, unless all players in the game have the same rating.

But, yes, there are probably bigger problems in terms of inflation.

I dont understand how those ratings work, but I definitely like what you did there. Cheers, Yonni.

Someone should work on a rating system that only factors in your last years worth of games

Durga wrote: ↑

Sun Jan 21, 2018 5:50 pm

Someone should work on a rating system that only factors in your last years worth of games

I can do this with GR. Maybe I will later.


Yonni, can I suggest that, at least at first, you make the weightings for all games identical to those for GR? It'll make it really easy to compare the two.

But as elo-based system, gr is not meant for short term evaluations - it is a long term rating which aims to your real rating ("in the limit").

Some months ago, a bug due to crashed games forced the vdip ranking to be limited to the games of the last couple of years only, and it was a real mess!

Yes, GR over the past year will mostly just demonstrate who played the most. Year over year GR improvement might be a more sane metric, but even then, all increases aren't created equal.

I thought about it more. Last 20ish FP games is a better metric

Ok nvm I've been told it isn't possible

That's a great idea ghug - do you know weighting we use for GR?

A ratings system that only factors in your last year of games is going to be very unstable.

Thanks for taking the lead on this, Yonni! Very much appreciated :)

It's really better to call GR "Elo inspired" rather than "Elo based". Many of the features that make Elo effective are not present (for better or worse) in GR.

All of the weightings are in Variantfile.csv in the git repo I posted. Second to last column, I think they're the in verses of the actual numbers you want (gunboat is .25 rather than 4).

Thanks, and what's the weighting for the 1v1 maps and for known world?

1v1 is weighted 0. It's really a separate game, they're all unranked, and it's two-part, so I just implemented Elo for it.