The Ghost Rating system seems to have very different philosophical principles than the Elo system

We know this. What is your point? :)

Wouldn't the actual Elo system be more accurate?

Maybe, if it were not impossible to implement for this case.

But I am sure no one will complain if you tried to do it.

All games are played with 1/17.5 of the players' collective points and distributed either like a WTA or PPSC depending on the game being played. The key is the players have no say in how much they contribute so you get a more even balance of play than with points from the site. And if you run low, you don't get refilled like you do on the site.

@ivo: it's not impossible... it's actually pretty easy. I'll post the details in a bit.

@Draug: yes, I understand how it currently works. You've done a good job of describing the net effect of the system. The point that I am making is that this is not like the Elo system at all and I think that the Elo system would be more accurate and maybe what he originally intended.

The big difference between Chess and Diplomacy is the number of players.

For Chess, the higher ranked player should have a significant advantage over the lower ranked player. The higher ranked player should almost always win if the rank difference is significant and accurate.

This is not true in Diplomacy as the higher ranked player can easily be ganged up upon. So exponentially scaling things wouldn't seem to make as much sense.

@Alderian, that's a good point. Although, anonymous games would prevent people for meta-gaming by rating. Of course, results in Diplomacy is a bit more random than chess, but this can be accounted for by dividing by a larger value before the exponential scaling. I think a major issue is the huge volatility factor. In Elo, the V factor is less for players with more games/higher ratings since it's likely that their ratings are already quite accurate and reflect their ability.

Also, results aren't necessarily that random...
userID=6454 TheMasterGamer (103 games, 51 wins)
I think there are some other examples

yebellz - this is very interesting... can you explain the 'volatility factor'? what does that mean? maybe I simply dont understand what the Elo system measures that the Ghost system does not... my understanding of Elo was that it would reflect players of varying ranks playing each other in such a way that wins against higher ranked players would be worth more than wins against lower ranked players.

but I do agree that the current ghost system has some short comings in that once you get VERY high, gaining points from a game requires you to win and only win... or get 2way draws. which is the situation MM faces.

I very much like the fact that the GR doesn't slowly creep up over time, like the points system. If you get lucky and win a few games and your GR gets pretty high, your GR should then go down over time if you can't keep up with the wins.

Both Ghost and Elo use the same basic formula:
New Rating = Old Rating + Volatility * (Result - Expected Result)
They just differ in how they calculate expected result and what they set for the volatility. In Elo, it's usually just a fixed, small value, whereas in GR it's the combined rating of all of the players divided by 17.5. This means that in high ranked games, players will get very large swings in rating, allowing for a greater possibility of a player jumping way ahead. With such a huge rating gap, MM would face the same problems even with Elo. Actually it would be worse, but it would be so much more unlikely for a player to ever create such an unrealistic gap in an Elo system. The GR system has allowed MM to take such a huge lead, but with the huge volatility at the top, large leads aren't very meaningful.

GR may be zero-sum to prevent inflation, but there's still huge volatility at the top, which does make much sense. The volatility shouldn't serve to pull people from the bottom if they don't keep up. It's the (Result - Expected Result) that should enforce the heavily weighted expected outcomes amongst uneven opposition.

http://www.diplom.org/Zine/S1998R/Nichols/ratings2.html

I thought when a new person registered they were given a default of 100. Wouldnt this cause inflation?

100 ghost rating i mean.

This is nothing people don’t know (I’ve told people myself before), but there were good reasons for doing what I did.

“Wouldn't the actual Elo system be more accurate?”

No, for two reasons:

1. The small number of games people play is a serious problem. Madmarx’s skill level would still be higher than his Ghost-Rating would indicate if I didn’t put in the scaling factor.

Originally the system didn’t have this, but it made a mockery of the top places in the system: MM and TMG were swapping places not based on who had played better that month but based on who had played *more* that month. One of the most important features of G-R is that it doesn’t matter if you’ve only played 30 games, you can still be highly rated.

Of course, you could have a steadily reducing volatility, but that would either introduce non-zero sum issues (if done individually) or defeat the point of the system and leave players unable to advance much against experienced players (if done on a by-game basis)

2. The rationale behind the Elo Expected result formula doesn’t work either. Elo argued that, in chess, if you play better than the other guy, you will win, and if you play worse than the other guy, you will lose. He modelled the standard of play of each player as a Normal distribution, and then looked at the probability that one player played better than the other.

In diplomacy this argument doesn’t work. Playing better than everyone else only gives you a better chance of winning, because a less good player might have the game more or less thrown to him by a real fool in a manner that you cannot do anything about.

Therefore I designed a rating system whereby people’s ratings reflect the relative average scores, and use ratios to calculate the expected result

No offense to TGM or anyone else, but you have to be totally insane to actually care who has the higher Ghost Rating. I looked at the list, and I saw two meta-gamers on the list ahead of me, who only drew because they never stabbed each other. I would report them, but i haven't seen them in a while.

“I thought when a new person registered they were given a default of 100. Wouldnt this cause inflation?”

No, because the average rating remains the same, and inflation is an increase in average rating.

In fact it is slightly more complicated than that, because it is an increase in the average rating of active players. Because bad players are more likely to leave, we have it that the system, from an initial situation, inflates slightly (I think it is by around 10%). However, the average rating then becomes stable, because there is a pretty constant value for average rating of all players: average rating of players who stick around.

This means that the rating system doesn’t inflate after a while.

"No offense to TGM or anyone else, but you have to be totally insane to actually care who has the higher Ghost Rating. I looked at the list, and I saw two meta-gamers on the list ahead of me, who only drew because they never stabbed each other. I would report them, but i haven't seen them in a while."

If players are inactive, the drop off the list. There is also the option to remove games between some combination of players too, however the way the banned players lists are kept makes that rather tricky to use effectively.

"but I do agree that the current ghost system has some short comings in that once you get VERY high, gaining points from a game requires you to win and only win... or get 2way draws. which is the situation MM faces."

This is true, but in order to get very high, you need to be able to get those results. Also, I don't think MM would be two impressed himself if he got a draw with a group of new players, anyway.

@killer - Report them to the mods and if the mods ban them, they'll come off the list. But if you don't report them then we'll just have to assume you are being bitter cause they wouldn't work with you and your caustic attitude.

I knew another player like you once. Two actually. Both starting acting more sane in their games and they started getting better results. I very much enjoy playing with one of them now where I used to play him just to see his rants.

A new player could also in principle cause some degree of perceived deflation. Suppose a very skilled player joined the site and quickly rose in the Ghost ranks by obtaining successes against top players. Since it is zero-sum, his rise is at the expense of others and lowers the ratings of other good players. However, this situation is quite unlikely, and I'm sure we have much more cases of bad players leaving, which adds a small amount of inflation.

Either way, with steady demographics, it all will settle down. Individual players really have no perceivable effect.

All good points, Ghost. Outcome in diplomacy is much more uncertain than in chess even with good players, however this could be accounted in the dividing constant pre-exponential scaling. In the article that I linked they divide each rating by 500 before the exponentially scaling.

What's a bit strange about using a simple ratio (no exponential scaling) is that basically points become less meaningful as they increase. A player with an 800 rating playing against 400 rated opponents has the same expected outcome as an 80 rated player against 40 rated opponents, despite there being a 400 point gap vs a 40 point gap in the different cases.

True... which is in fact why the variance changes.

In fact, having a fresh look at it this afternoon, you could put:

R=ln(G) / c
to convert from Ghost-Rating (G) to Elo-Rating (R)

So the difference is really rather cosmetic.

My thoughts on rating systems:
- The standard 'pot' rating system the site uses has severe flaws and has been gamed before.

- The Ghost Rating system is based loosely on the Elo, has been and will be continually modified (until TGM gets sick of modifying it) to better reflect the players style and skill. It is easily the better of the two systems used.

- If you have a better way to do it, please let a developer that has time know, and you can setup rankings for your system. No one will have an issue having 4 or 5 different ranks because one system will never, truly, demonstrate your exact skill compared to others, especially when it is a 7 (or more/less in some mods) player game. The best we can do is get several 'close' rankings to see how we fit compared to others.

So, if you want to start the "hebellz 'closer to elo' rankings", either start coding, or explain it to someone who is willing to code it for you.

@Ghost, it's still not quite that easy to convert from one to the other due to the way the variance/volatility is set. I don't see a huge issue in introducing a little bit of non-zero sum behavior into the rating system.

@fortknox, I would be happy to work on some code to implement it. I don't think it would actually require much modifications from the Ghost rating code, if Ghost would be willing to provide me with that. I heard that it was written in Perl, which I'm not an expert in, but I think could manage.

The link that I posted earlier roughly describes what I want to do
http://www.diplom.org/Zine/S1998R/Nichols/ratings2.html
but I'll adjust the parameters and simplify the way that the variance is set. I can flesh out my ideas a bit and write up a short description. Give me some time though.

Would I be able to get someone to run the code for me? I guess it would have to interface with the database somehow. Or does it just crawl the website?

"@Ghost, it's still not quite that easy to convert from one to the other due to the way the variance/volatility is set. I don't see a huge issue in introducing a little bit of non-zero sum behavior into the rating system. "

It depends how it acts. I really like the way that you can compare over time with a non-inflating/deflating system.