ELO calculation

[maknud]

I thought for sure there was a thread on this somewhere, but I couldn't seem to find it. Is the ELO calculation still as specified on RavenBlack's rules, games.ravenblack.net/rules/1/elo.html ? So killing is better than letting them surrender?

[awall]

Yup, but only barely. IIRC, 90% of the ELO is earned for a win, and the last 10% comes with a kill.

[Rycchus]

Dubber made a neat little chart of it: dubber.org/rb/neato.htm

[Dubber]

I think someone showed that my ELO chart (which I cribbed from Garfield) used the original maths, which were modified by RavenBlack to remove the win for negative ELO problem identified a few months after implementation.

[maknud]

That's what I thought I remembered reading. So the formula listed on RB is wrong, then. Anyone know what it actually is?

[Rycchus]

Read that thread earlier today. First I'd heard of the modification, tbh. Someone mind explaining to me what the negative ELO problem was?

[maknud]

At least according to the stated formula, ELO tracks the expected value of some score you can get in a single game. This score has value 1 for killing the opponent and 0.9 for just winning. So if you're really good against someone not very good, your ELO drops (rightfully so!) when your opponent manages to keep you from getting a 1, i.e. surrenders and makes your score 0.9. If everyone knew exactly how good they were, and if ELO worked perfectly, then it should stabilize at the expert having an expected score of 0.9 in a game, not 1.0, as the opponent could just surrender instead of playing.

Which in turn would effectively cap ELO since if someone has too much you can just play them and surrender 'til they're back down to 0.9 land.

[Rycchus]

Yes, the 0.9 thing has been heavily criticised because RavenBlack took a perfectly good working ELO system and modified it into something that wasn't actually ELO.

I'm not sure what "If everyone knew exactly how good they were" is meant to imply.

I think the mod RavenBlack did (now that I remember what the problem was) was just to set any negative-Elo'd wins back to zero, i.e. if I surrender to someone who would have got -2 for the surrender, it's just 0 instead. But I could be wrong.

[maknud]

It is still ELO. ELO doesn't measure likelihood of winning, it measures expected score - which in the extreme case of 1.0 for winning, 0.0 for losing, and no other possibilities is equivalent to likelihood of winning. You can define the score however you like, you just get the unwanted effects mentioned if you happen to define it as not-a-perfect-score if your opponent surrenders rather than dying.

The "if everyone knew..." is saying that if you know for a fact your expected score sans surrendering out of hand against Toyotami is 0.05, then your "best" (in terms of ELO) action is to immediately surrender for a score of 0.1 instead of carrying on any maybe getting trapped by Toyotami's superior play into not being able to surrender and ending up with a 0.0.

[Rycchus]

Well I'm not sure it is ELO for exactly the reasons above. If someone's several hundred below you, they might be dumb enough (from your perspective) for you to be able to kill them, but they might equally be dumb enough to para-surrender or something. This isn't a case of "ha, so-and-so's not REALLY at that super-high ELO coz he couldn't kill dumb-newbie-guy", so when so-and-so's ELO is modified negatively to account for this I don't think that's a true ELO calculation. This is basically because I don't believe that even a truly experience player has enough control over the die/surrender outcome to be able to say that their expected score is 1.0 (of course, it won't ever be perfectly 1.0, but close enough to 1 to enough significant figures that the difference is negligable).

Perhaps you think my "it's not ELO" statement is too strong but it seems that we're basically in agreement, we're just expressing it differently.

[maknud]

Yeah, here's what I think our sole difference is:

-it is possible to assign a score for each game with values {0.0, 0.1, 0.5, 0.9, 1.0} whose expectation can be approximately tracked by dynamically assigning each player a number

-I call this number one possible correct implementation of ELO, which does not do what we expect or want something called "ELO" to do

-You note that this number does not do what we expect or want something called "ELO" to do, and so say this number is not a correct implementation of ELO

The quibble is tiny and only applies to how we think the term "ELO" should be used, not any of the underlying structure. So... anyone know how the number labeled "ELO" on RB is actually calculated? Is it as Rycchus remembers, making the maximum loss 0 points? If so, what effect does this ad hoc and arbitrary change to the mathematics imply in the limit?

Yay maths!

[xade]

ha! You guys, quibbling over the term ELO.

Did you know!

ELO actually wrote Xanadu. I did not know that.

I wonder if they had an 0.9 for surrenders when they worked that out...

[Rycchus]

Searched Warlocks University Yahoo group and couldn't find any references anywhere; I presume the discussion with RavenBlack (assuming it was public at all) was on another Yahoo group - but the others are members-only and I couldn't be bothered to sign up to a defunct group just to check this out. Dubber?

[Dubber]

ravenblackgames:
#441 Aug 15, 2004 Yaron to rajtheguru:

Re: [ravenblackgames] small ELO ranking flaw?

Rajiv,

I believe Raven Black has fixed this so that any positive score for losing (or negative score for winning) is replaced with zero.

Incidentally, the maximum score to be gained from a loss (before the fix) was only 2.5 points, and even that at score differences approaching infinity. You are correct in that 366 points is the threshold to gain anything.

Yaron


Rajiv <rajtheguru> wrote:

I just read how the ELO system ranking worked, and i think i found a small/minor flaw...

ok, the scenario is so- a new player joins.... with ELO 1500,

and plays a ladder/friendly with another player whose elo is 366points higher (ie ABOVE 1866)

in their inexperience they surrender first turn, however this would mean their ELO would go UP!

if they had played the top play (which at present time is prioli ELO 1972) a posotive of 472points.

they would gain 10ELO points and would be on 1510.

i know its a small difference, but as more and more players enter inevitable the ELO difference would become greater, as well as the consequences.....

Rajiv

copy paste error fixed

[Dubber]

btw - people ask about secondary sorting a lot, too:

ravenblackgames #394 April 2, 2004

e: [ravenblackgames] Elo in the Rankings

>I know it's not up there yet, but what better time to make requests then
>during development? I'm wondered where Elo will fall in with the rankings
>when it does get there. I think it would fit best as a second or third
>priority, after either Ladder/Melee scores, or kills/deaths. I think it
>would fit best in the former, making it so that rankings of equal
>ladder/melee points are then decided by the highest elo. Thoughts?

Hardly worth it - that would essentially only sort amongst people who have
only one or two ladder points, and the current sorting by win/loss ratio
will be pretty much the same. However, the Elo rating is now shown on the
player list, and the list can be sorted by score, melee score, or elo, at
the viewer's preference. (Secondary sorting remains based on win/loss ratio.)

--RavenBlack

blockquote error fixed