KGS ranking revisited

[hyperpape]

But you're not doing that work and you're ignoring the fact that KGS is based on one of the better models of ratings we have.

[Javaness2]

Why can't everyone just pick the rank they want to have, and play at that rank?

[Sverre]

Why can't everyone just pick the rank they want to have, and play at that rank?

Because I want a 7d rank, but am unable to play at the appropriate strength

[jts]

If a 1.1-dan plays a 1.9-dan, there are at least two reasonable options:

The pairing system can say both are 1-dan, so the game is even (komi 6.5)
The pairing system can say rating difference = 0.8, so the game is uneven (komi 0.5)

I gather KGS uses the first option. I play mostly on IGS, where I believe the second option is implemented. IGS also has a command to predict win/loss probability against any opponent, based on exact ratings and handicap and komi, but the calculation has bugs and probably will never be fixed. Every system has problems

You can play improperly handicapped games on kgs. It works exactly as you describe.

[RobertJasiek]

But you're not doing that work

I know. I cannot do all work for everything.

and you're ignoring the fact that KGS is based on one of the better models of ratings we have.

It is a frequently used excuse. As long as the system has major drawbacks, being based on one of the (as you claim) better models is by far insufficient.

[Charlie]

Assuming that a rating system is designed in such a way that its quirks are less significant than the natural variations in a player's strength, I believe that it is wrong for that rating system to punish a 1.9 dan for losing to a 1.1 dan.

Within a band, players should play even games and wins and losses should have a symmetrical effect on player's ratings.

The only possible argument against this is that the 0.8-rank difference between 1.1 and 1.9 does, in fact, represent a significant deviation from 50% win-probability - one that is more significant than the natural variations in the player'
s strengths. If this is the case, then the assumption stated in my first paragraph is not true and a solution must be sought.

The simplest solution is obvious: lower the band-width of the ranks.

At the very least, KGS should show the 1.9 dan that they are the "favored" player. It doesn't have to be as blatant as a message-box saying they are expected to win with a 67% probability. Simply including the word "favored" under the player's name in the side-bar would suffice.

Another thought crossed my mind. (It's rare, I know, but it crossed at speed and didn't look back.)

If you win or lose a game against a player and that player subsequently drops in rank, does that win or loss "drag you down" or simply become less significant?

[Mef]

Assuming that a rating system is designed in such a way that its quirks are less significant than the natural variations in a player's strength, I believe that it is wrong for that rating system to punish a 1.9 dan for losing to a 1.1 dan.

Within a band, players should play even games and wins and losses should have a symmetrical effect on player's ratings.

The only possible argument against this is that the 0.8-rank difference between 1.1 and 1.9 does, in fact, represent a significant deviation from 50% win-probability - one that is more significant than the natural variations in the player'
s strengths. If this is the case, then the assumption stated in my first paragraph is not true and a solution must be sought.

The simplest solution is obvious: lower the band-width of the ranks.

At the very least, KGS should show the 1.9 dan that they are the "favored" player. It doesn't have to be as blatant as a message-box saying they are expected to win with a 67% probability. Simply including the word "favored" under the player's name in the side-bar would suffice.

Another thought crossed my mind. (It's rare, I know, but it crossed at speed and didn't look back.)

If you win or lose a game against a player and that player subsequently drops in rank, does that win or loss "drag you down" or simply become less significant?

It is not a punishment per se, it is that the 1.9d player is not taking a symmetrical "risk" so to speak. The 1.9d gets the benefit of an under handicapped game that they are expected to win with higher probability. The effect is that they potentially see a lower "reward". Of course in this context the whole idea of "punishment" vs. "reward" is not really right in the first place...What you actually have is two predicted variables (ratings of the players) that will change as new information becomes available. Instead of thinking it as "your rating goes down" think of it as "the server improves its estimation of your actual rating, which is never truly known".

When you think of it this way, it makes sense why the system works the way it does: If you are trying to predict what will happen, and your predictions are correct, there is little need to alter your prediction model. If something highly unexpected occurs, that suggests you need a proportionally large change. If you have something you are highly uncertain about (an even game) then the new information can be used to help you predict better in the future.

Likewise if you beat an opponent and their rating goes down, it is not that they will drag you down, however they will no longer boost you as high as it originally did. This comes back to the same principle - If you beat someone who is rated at 5k, and later it is found that 5k was in error and a new, better estimate of their rating is 7k, the win will now be counted as if it was against a 7k. Aside from increasing accuracy of the system as a whole, this has a number of pleasant side effects for people who are still working toward good estimates of their rating (unintentional sandbaggers, people who accidentally get rated much too high when first starting out, or anyone in between). Imagine someone logs on to KGS unsure of their rating but they play at 5k strength...they play a 10k and win, get a 10k? rating. Maybe they have a connection issue on their next game and lose by time to a 9k, making this person a solid 10k. As they proceed to go on a tear and beat their next 8 opponents while their rank goes 8k, 7k, 6k...The people who helped this person get to their correct rating shouldn't have their losses counted as against a 10k, because we now know this person is a 5k. Instead they are all counted as losses to a 5k, because that's what they were.

[bleep]

I never realised what a contentious subject this could be.

I just wanted to chime in as a happy customer - KGS ranking seems to make perfect sense to me. I play roughly a game a day, and if I win, I go up a little, if I lose, I go down a little. A streak in either direction sees the graph respond as I would expect.

[Ortho]

I am not sure how I feel about the rating system. I was at 10k when I made some improvements and it took me posting a really good record at 10k over 25 games or so to change to 9k.

From the time I went from 10k to 9k I had to go 39-25 (so 64 games) before my rating changed to 8k.

I am now 27-17 as 8k and the graph shows me at about halfway to 7k, so I guess if I keep going at the same pace it'll be 50-75 games before my ranking changes. I feel pretty clearly that I'm a stone, possibly 2, almost certainly not more, underranked. It does seem like a lot of games and I do feel like most of my games are too easy, but if I am only ranking up like this once and it will solidify sometime or other, I guess it's not a big deal. It certainly doesn't lag as much behind reality as my over-the-board rating.

[hyperpape]

27-17 is a 61% winning percentage, which I believe is actually worse than you'd expect if you were a 7 kyu. You are probably not significantly underrated.

(There are a few complications related to handicapping within ranks and all that, but I've performed a subtle analysis that I can't include here crossed my fingers and hoped that won't matter).

[Ortho]

27-17 is a 61% winning percentage, which I believe is actually worse than you'd expect if you were a 7 kyu. You are probably not significantly underrated.

Perhaps you're right. Only one way to find out.