More on A Curious Case Study in KGS Ranks

[Bantari]

I see an immediate success of a win. Its a win. Not sure why this is not enough.

Because... KGS does not reward a win properly as a win for frequently playing players. KGS only rewards a win properly for frequently playing players if they invest so much successive effort and time in creating enough wins and a sufficiently high winning percentage that they have to have that great amount of effort and time available, because they need not invest any effort and time in a job or other essential activities of life. A frequently playing player can make himself a slave to the KGS rating system and devote all his life to fitting its requirements, or he has no good chances of reaching the rank, at which he meets a distribution of opponents against whom he would win ca. 50% even real world games.

"A win is a win" is enough only for infrequently playing players, because they need not become the slaves of the KGS rating system in order to reach the rank, at which they meet a distribution of opponents against whom they would win ca. 50% even real world games.

On-line rating is good for getting you to reach the distribution of opponents against which you win ca.50% of your online games. No more, and no less. No system can solve that problem for you because the problem is inherent to personal variation between online and real-life play most people display not to system calibration. It has nothing to do with frequent play or non-frequent play.

[ez4u]

It sounds like the ideal rating system for Robert is L19's: Type whatever number in the box makes you happy

I love it! But to be honest, I think this is true for a lot more of us than just RJ.

[Hong Ny]

It sounds like the ideal rating system for Robert is L19's: Type whatever number in the box makes you happy

I love it! But to be honest, I think this is true for a lot more of us than just RJ.

Try OGS, you can change your ranking yourself.

[Mef]

It sounds like the ideal rating system for Robert is L19's: Type whatever number in the box makes you happy

I love it! But to be honest, I think this is true for a lot more of us than just RJ.

Indeed, and when I thought about this later I realize how much so many people do this simply by the nature of having so many ranking systems out there. If a person thinks that they are a 5k, and they have 6 different ratings in 6 different places ranging from 10k-1d....they will pick the places that puts them as 5k and qualify their statement ("Oh I'm 5k AGA/IGS/KGS/Wherever gives them the rank they think is right).

[RobertJasiek]

JFTR, self-setting a ranking is not what I want to see in a rating system.

[RobertJasiek]

It has nothing to do with frequent play or non-frequent play.

You still have not understood the problem.

[Aidoneus]

It has nothing to do with frequent play or non-frequent play.

You still have not understood the problem.

I looked up the KGS rating system to try and understand this thread. I learned that rank adjustments are based on ranked games played in the last 180 days and the games are weighted by how long ago they were played. (I assume newer games carry more weight, right?) So, if I understand you, someone who plays very frequently will need to do well consistently over six months. Whereas someone who doesn't, say someone who goes off for several months and then comes back to play again, will have to obtain consistently better results for a much shorter period (less games). Yes?

If this is how the system works, I can think of a plausible explanation for such a design choice. Say someone abandons Go for three months or more, and now they play well below their former strength. In this case, their rank will drop rather quickly to bring it into alignment with their true current strength. On the other hand, a player who spends three months studying and playing exclusively over-the-board games may be significantly stronger when they return to online play. Again, their rank will adjust faster for only considering their recent games.

Of course, what I really don't understand is why there isn't a more transparent finer-grained system. I'm thinking of something like the Elo system for chess, where each class (think rank) varies over 200 rating points. Chess players can readily see how close they are to advancing or declining in class, especially as the number of rating points gained or lost from each game is simple to calculate.

Please be gentle with me if I have simply confirmed my KGS ignorance!

[Boidhre]

I looked up the KGS rating system to try and understand this thread. I learned that rank adjustments are based on ranked games played in the last 180 days and the games are weighted by how long ago they were played. (I assume newer games carry more weight, right?) So, if I understand you, someone who plays very frequently will need to do well consistently over six months. Whereas someone who doesn't, say someone who goes off for several months and then comes back to play again, will have to obtain consistently better results for a much shorter period (less games). Yes?

The flip-side that often seems to be forgotten is that instability that allows your rank to rise quickly when you're playing well and playing infrequently also means it drops quickly if you hit a losing streak. The faster adjusting of rank works in both directions, the person who plays a lot and gains a rank will have a much more resilient rating than the person who is playing little and whose rank is tied to relatively few past results.

[RBerenguel]

It has nothing to do with frequent play or non-frequent play.

You still have not understood the problem.

I looked up the KGS rating system to try and understand this thread. I learned that rank adjustments are based on ranked games played in the last 180 days and the games are weighted by how long ago they were played. (I assume newer games carry more weight, right?) So, if I understand you, someone who plays very frequently will need to do well consistently over six months. Whereas someone who doesn't, say someone who goes off for several months and then comes back to play again, will have to obtain consistently better results for a much shorter period (less games). Yes?

If this is how the system works, I can think of a plausible explanation for such a design choice. Say someone abandons Go for three months or more, and now they play well below their former strength. In this case, their rank will drop rather quickly to bring it into alignment with their true current strength. On the other hand, a player who spends three months studying and playing exclusively over-the-board games may be significantly stronger when they return to online play. Again, their rank will adjust faster for only considering their recent games.

Of course, what I really don't understand is why there isn't a more transparent finer-grained system. I'm thinking of something like the Elo system for chess, where each class (think rank) varies over 200 rating points. Chess players can readily see how close they are to advancing or declining in class, especially as the number of rating points gained or lost from each game is simple to calculate.

Please be gentle with me if I have simply confirmed my KGS ignorance!

That's more or less (too long since I checked EGF ad ELO ways of computing formulas to be sure if it's very similar or just similar, I don't remember ELO having a certainty factor, but it had a new player factor) how the EGF rankings work.

[skydyr]

Of course, what I really don't understand is why there isn't a more transparent finer-grained system. I'm thinking of something like the Elo system for chess, where each class (think rank) varies over 200 rating points. Chess players can readily see how close they are to advancing or declining in class, especially as the number of rating points gained or lost from each game is simple to calculate.

Please be gentle with me if I have simply confirmed my KGS ignorance!

As I understand it, the KGS ranking system does keep track of rank or rating on a very fine-grained level internally. As a design decision, however, it was decided that games between 2 primary ranks, like 3k and 4k, should always be handicapped at the same level (1 stone in this case) regardless of where within the rank the two players are. This was done so that if you start a game with a player one stone stronger as displayed, it's always a one stone game, and won't surprise you with a 2 stone or even game because the rating difference was actually 0.5 or 1.75. I believe the rating system does take this into account when it calculates the win probability, though.

[SmoothOper]

Of course, what I really don't understand is why there isn't a more transparent finer-grained system. I'm thinking of something like the Elo system for chess, where each class (think rank) varies over 200 rating points. Chess players can readily see how close they are to advancing or declining in class, especially as the number of rating points gained or lost from each game is simple to calculate.

Please be gentle with me if I have simply confirmed my KGS ignorance!

As I understand it, the KGS ranking system does keep track of rank or rating on a very fine-grained level internally. As a design decision, however, it was decided that games between 2 primary ranks, like 3k and 4k, should always be handicapped at the same level (1 stone in this case) regardless of where within the rank the two players are. This was done so that if you start a game with a player one stone stronger as displayed, it's always a one stone game, and won't surprise you with a 2 stone or even game because the rating difference was actually 0.5 or 1.75. I believe the rating system does take this into account when it calculates the win probability, though.

I think this is what makes IGS and KGS feel so artificial and rigid, the players strengths start to cluster at the stone boundaries, looking for that one extra win, then when they accumulate enough wins to go up a stone, they end up giving a handicap to essentially equal players, and can't maintain the win rate. With Tygem you essentially change ranks when you are significantly different from the mean of the rank, which is important because the emphasis isn't placed on the boundary, but the mean.

[skydyr]

I think this is what makes IGS and KGS feel so artificial and rigid, the players strengths start to cluster at the stone boundaries, looking for that one extra win, then when they accumulate enough wins to go up a stone, they end up giving a handicap to essentially equal players, and can't maintain the win rate.

When the rating system calculates the win probability, it takes this into account, so if you are just barely above a boundary playing someone just barely below, the win probability as calculated will be skewed in favour of the slightly weaker player, because they got a free handicap stone without the rank difference. As a result, in a series of games, the slightly weaker player would need to maintain a win percentage much higher than 50/50 to maintain their rank, and the stronger player would need to maintain one much lower, maybe 35-65 or 40-60, depending on the actual rating difference.

[Aidoneus]

I think this is what makes IGS and KGS feel so artificial and rigid, the players strengths start to cluster at the stone boundaries, looking for that one extra win, then when they accumulate enough wins to go up a stone, they end up giving a handicap to essentially equal players, and can't maintain the win rate.

When the rating system calculates the win probability, it takes this into account, so if you are just barely above a boundary playing someone just barely below, the win probability as calculated will be skewed in favour of the slightly weaker player, because they got a free handicap stone without the rank difference. As a result, in a series of games, the slightly weaker player would need to maintain a win percentage much higher than 50/50 to maintain their rank, and the stronger player would need to maintain one much lower, maybe 35-65 or 40-60, depending on the actual rating difference.

Ah, thank you both for the explanations. Like I thought, the gains/losses from each game are not entirely transparent owing to some hidden variable(s) within each person's rank.

It also seems that having stone handicaps is both a blessing (for more competitive games) and a curse (for calculating rating changes).

[Mef]

I think this is what makes IGS and KGS feel so artificial and rigid, the players strengths start to cluster at the stone boundaries, looking for that one extra win, then when they accumulate enough wins to go up a stone, they end up giving a handicap to essentially equal players, and can't maintain the win rate. With Tygem you essentially change ranks when you are significantly different from the mean of the rank, which is important because the emphasis isn't placed on the boundary, but the mean.

If these clusters happened it would be a self correcting problem. If you are a person on the borderline of promoting and you are playing people who are also borderline to promote, then all you must do is have a >50% win rate. Likewise, if you are a person who is borderline promoted, but you are playing someone who is borderline about to promote (below you) at no komi, then you are expected to only win about 33% of those games. Even with a lower than 50% winrate your rating would keep going up.

[SmoothOper]

I think this is what makes IGS and KGS feel so artificial and rigid, the players strengths start to cluster at the stone boundaries, looking for that one extra win, then when they accumulate enough wins to go up a stone, they end up giving a handicap to essentially equal players, and can't maintain the win rate. With Tygem you essentially change ranks when you are significantly different from the mean of the rank, which is important because the emphasis isn't placed on the boundary, but the mean.

If these clusters happened it would be a self correcting problem. If you are a person on the borderline of promoting and you are playing people who are also borderline to promote, then all you must do is have a >50% win rate. Likewise, if you are a person who is borderline promoted, but you are playing someone who is borderline about to promote (below you) at no komi, then you are expected to only win about 33% of those games. Even with a lower than 50% winrate your rating would keep going up.

Self correcting like a yoyo? Anyway, playing someone equal and giving handicaps isn't what I consider a good game?

I also want to point out that the problem is exacerbated by the rank above, being clustered at that boundary, and suddenly playing with no handicap.