KGS Ranking adjustment?

[RobertJasiek]

You can "justify" any rating system (such as "the rating is a constant 1900") by saying "calculates the true rating according to the rating system". Bullshit! A good rating system can be recognised by producing numbers that are meaningful from external analysis and easily understood by human perception. I have never seen either for the KGS rating system. It is nothing but a black box that works in itself and justifies itself internally claiming to reveal eternal truth for the external observer, until an admin admits nonsense by readjusting some parameters.

[lodestone]

I don't think the peculiarity you noticed is due to the frequency at which the rank updates, I think it is due to the fact that it looks at all recent history for your previous opponents as it updates.

For instance, if I win a game against someone who is currently listed at 5kyu, but who is rapidly improving so that one week later they are up to 2kyu, the KGS rating system will treat my win as though it was against a 2kyu player instead of a 5kyu player.

That means that as my vanquished opponent's rank is improving, mine is also drifting upward along with it. That is most of the reason that rank on that server tends to drift around even when you haven't played games recently.

[Harleqin]

Bullshit yourself. You just can't separate the noise from the signal mentally, and seem unable to grasp what the signal itself is saying.

KGS' ranking algorithm certainly is far from perfect, but at least it is a more plausible effort to actually estimate strength than all ELO-based systems, especially given the paucity of data. If you want to gamble for points, then it may seem off to you, but gambling for points is, in the end, incompatible with actual strength estimation.

If you don't like the mechanisms of the KGS rank meta-game, then don't play it. Ignore the rank graphs and just be content that it gives you reasonably strong opponents.

[RobertJasiek]

KGS' ranking algorithm certainly is far from perfect, but at least it is a more plausible effort to actually estimate strength than all ELO-based systems

I find the EGF rating system (ELO) to be (far from perfect but) very much more meaningful than the KGS rating system.

just be content that it gives you reasonably strong opponents.

Nice theory, but in practice on average it does not. Among same-ranked opponents, I get players that play between ca. 1k and 6d real world ranks.

A simple rating system, such as +0.1 ranks for a win, +0 for a jigo, -0.1 ranks for a loss, would be much better.

[jlt]

Among same-ranked opponents, I get players that play between ca. 1k and 6d real world ranks.

People don't always play at the same level on the internet. The 1k wanted to play a relaxing game after a hard day's work and his baby started crying during a complex fight. The 6d was more concentrated because he wanted to recover the rank he lost the day before due to a disconnection problem.

A simple rating system, such as +0.1 ranks for a win, +0 for a jigo, -0.1 ranks for a loss, would be much better.

That's more or less the IGS ranking system. Indeed it works reasonably well and is easy to understand, the only drawback is that you need to play 20 games before you get a firm rank (without "?" at the end).

[Harleqin]

That's what I meant with gambling for points.

Such a system only works reasonably well, if

- the participants are paired randomly (otherwise subpopulations emerge that have different points distributions)
- each participant continually plays a lot of games (otherwise the noise of temporarily inactive players suddenly appearing with a different actual strength leads to different frustrations)
- the mechanisms for getting new points into the system in order to represent actual strength growth match the actual strength developments (otherwise there will be rating drifts and waves, but this is totally unmeasurable anyway)

It is very implausible that these assumptions are even remotely approximated.

So, all you get is gambling for points. That's simple and good for the endorphine rush of a simple mind, but that's all it is.

[Uberdude]

From some testing a while ago (few games in history, previous opponent turning rank on and off) I concluded the KGS rank recalculation is done of the order of every 10 seconds.

[RobertJasiek]

Uberdude, why then are rating graphs updated only once a day?

[jlt]

Updating graphs every 10 seconds would use too much computing resources.

[RobertJasiek]

Updating graphs (copying an already calculated rating) must require much less computing power than updating ratings (calculating rating).

[jlt]

Don't know.

15kB (size of a rank graph) * 20 languages * 30000 users = 9 gigabytes

do you want to update that every 10 seconds?

[RobertJasiek]

Are you suggesting that rank graphs would be distributed as pictures instead of rating numbers?! Does CGoban not generate the rank graph picture from only receiving rating numbers only when the user opens it?!

[jlt]

Actually I wrongly thought that rank graphs were stored as images. The rank graph in english of a given user can be visualized on

Code: Select all

http://gokgs.com/servlet/graph/nicknameoftheuser-en_US.png

I didn't notice the URL contains "servlet", which suggests that the image is generated at each request.

[Uberdude]

Uberdude, why then are rating graphs updated only once a day?

You'd have to ask wms, but I presume a combination of pragmatism, necessity, and laziness.

[jlt]

Does anybody else feel that KGS ranks have become weaker?

I looked at the table from a 2018 survey https://idex.github.io/go-rank-survey/g ... mean-ranks and circled my current ranks.

Capture.JPG (99.58 KiB) Viewed 32006 times