I don’t really understand what you want to say.
Do you think 1.8k ± 0.8 should be displayed differently (like -2.8)? Or are you fine with that and you would like 1780 ± 61 to be removed?
Okay, I only just understood the current system or so I believe. It’s the same as the AGA system but with the -1.0, 1.0 interval collapsed, and flipped (but that doesn’t really matter, other than it’s nice to have share conventions)
but collapsing the -1.0, 1.0 interval means NOBODY EVER has a decimal rating matching their actual ranking. All kyus have a rating in the rank above and all dans have a rating in the rank below
edit: actually it’s not just 0.0 between dans and kyus; for dan ratings there’s an adjustment—you can be 0.3k but not 0.3d. There’s a missing interval but it’s assymetrical, being only [0, +1.0]. This is just so confusing
Anyway, I’m a weak 2k (?) so my rating should be 2.8k, which everybody can understand is close to 3k. How am I supposed to know 1.8k is close to 3k ?!?
1 Like
Personally I’d prefer ratings without missing intervals, like with the FFG rating system.
…
2k corresponds to the interval [-200;-100)
1k corresponds to the interval [-100;0)
1d corresponds to the interval [0;100)
2d corresponds to the interval [100;200)
…
However I also agree that displaying 2.8k in your case would be less confusing than displaying 1.8k.
2 Likes
That would be EGF ratings shifted downward by 2050 points.
I feel that choosing the boundary between 1k and 1d as the 0 point of the rating scale, attaches too much importance to that boundary.
Even without this, that boundary already seems to have almost mythical status. There seems to be a notion that everyone below it is basically a patzer, while everyone above it is basically an expert.
I feel this notion is unjustified. Or at least I wouldn’t put that boundary between 1k and 1d.
If I were to choose the boundary between patzers and experts, I would place it higher up, somewhere around 7d EGF (which is taken to be the boundary between amateurs and professionals).
Though if you’d ask professionals, they might place that boundary higher still. So I think this boundary is a rather subjective thing.
Instead of focusing on such a subjective boundary between patzers and experts, I’d prefer to use something more objective, similar tot the EGA handicap number used for golf in Europe (ranging from 0 to 54, where more skillful players have a lower handicap number).
In go it would be based on how much handicap one needs against top level play and lower tier players can be ranked by their respective required handicap against higher tier players.
Since we have AI, we could set the 0 level to level of the strongest AI running on a high number of playouts on some powerful hardware. Stronger professionals might then be about level 3, 1d EGF about level 10, 20k EGF about level 30 and a typical raw novice 7 year old might be about level 50.
Alternatively, 0 could be set to the average level of (say) the top 5 human go players in the world. Probably only AI would then have a (small) negative handicap number.
While I agree that the boundary is somewhat arbitrary (kyu players in some rating system may be dan in another), I think you are exaggerating a bit. I understand your idea however but I still think that the zero should be set between 1k and 1d because it’s easier to convert a rating into a kyu/dan rank. Can most people tell without hesitating which kyu rank corresponds to an EGF rating of 1217? I don’t think so.
This amounts to saying that top pros are 3k, weak pros are 6k, etc. but differs too much from the tradition to be acceptable. I like the old Japanese convention: a 1d is 9 stones weaker than a top pro. This is probably close to the EGF rating system.
I’m confused, isn’t all this adressed by the elo-like rating (1680, in my case) ?
the only problem I see in the present systems is that decimal kyu ranks are one-off compared to integer kyu ranks
I guess it depends on what you expect that number to mean.
Also with Elo ratings, the 0 point of the scale can be chosen arbitrarily. Really only the difference in Elo ratings has a meaning (mapping to winning probabilities). In chess, Elo ratings are historically anchored to a central value of 1500 which is supposed to correspond to the most common chess player (“middle” of the rank if you will). I think OGS just took over that convention from chess. But there would have been nothing wrong technically with setting the most common player level at 0 rating points.
So your Elo rating of 1680 means little more than that you’d probably win about 74% of even games against a “middle” OGS player (Elo rating 1500).
Relative Elo ratings are conventient to express such probabilities, but rather than having this benefit, @jlt prefers to have ratings that allow easy conversion (from the head) to traditional go ranks. Elo ratings don’t provide that convenience (no matter where you place the 0 Elo rating).
I couldn’t do that either. Indeed conversion between EGF ratings and ranks would be easier if all EGF ratings were shifted downward by 2050 points.
However, I’d expect some resistance against resulting negative ratings. It was one of the reasons why some people involved in the 2021 EGF rating system update resisted to lowering the EGF rank floor from 20k to 30k (resulting in negative EGF ratings for players weaker than 21.0k). I think they were worried that negative ratings might result in a negative social stigma.
1 Like
I think if you’re 6.97k it will display as 7k and 7.0k, right? In your proposal it would be what, 7k and 8.0k? We might need some unusual rounding rules to completely fix it.
Are you sure that your understanding is correct? If it is, then I completely agree with you, it’s entirely confusing (not to say outright wrong) to write 1.8k for “weak 2k, close to 3k”. But are you really sure that this 1.8k means “weak 2k” and not “weak 1k, rounded to 2k”? I’m genuinely in doubt now.
I think there is a difference to be made between a notation using k and d, and a notation using + and -.
- When we use k and d, the scale shouldn’t care about a gap between 1k and 1d, and should instead be focused on making the values meaningful relating to kyus and dans; so 1.8k should mean “weak 1k, close to 2k”, and 1.8d should mean “strong 1d, close to 2d”. I have always thought that was what the OGS displays meant but now with @pokk97 's comment I am in doubt.
- When we use + and -, the scale should make sense as a numeric scale, so there shouldn’t be any gap near 0. This is the way the FFG does it, where [-100, -1] means 1k, and [0, 99] means 1d.
Yes, I confirm he is correct, see the screenshot I posted earlier:
Compare with this
Wouldn’t 6.03k display as 7k and 6.0k? The problem is kind of unavoidable.
Edit: actually I realize the bottom end doesn’t matter since the entire rank doesn’t match. 6.00k-6.9k doesn’t match 7k so it’s still worse than the rare instance of 8.0k not matching 7k
We might as well just calculate the rank off the rounded result rather than sourcing both to the actual value. 7.97k rounded to 8.0k should display as 8k.
1 Like
When displaying k/d ranks with decimals, I would opt for 2.0k to mean “middle” 2k, so the range of 2k would be <2.5k,1.5k>, where 2.5k would be the weaker direction and 1.5k would be the stronger direction.
Similarly, 2.0d would mean “middle” 2d, so the range of 2d would be <1.5d, 2.5d>, where 1.5d would be the weaker direction and 2.5d would be the stronger direction.
Decimal ranks in the range <0.5k, 0.5d> would not exist.
2 Likes
to decrease confusion <1.51d, 2.49d> would be better,
.50 should never be displayed
With the notation <1.5d, 2.5d> I meant to say from 1.5000…1 to 2.4999…9.
No, there is no issue here. In a system based on flooring numbers (where 6.97k is a 6k rating) you should simply be consistent and floor the decimal ratings as well; 6.97k should display as 6.9k
A meaningful numeric ranking is why we have the ELO system. The kyu/dan rating is a display issue and as such doesn’t need to make sense numerically, although I agree displaying 1.3k / 2.1d is much preferable to -1.3 and +2.1 and this removes any possible confusion
Whether the border between ranks is 2.499 or 2.999 is entirely arbitrary. There’s nothing wrong with using flooring; having a a rank of 2k with a 2.7k rating really isn’t confusing
One big argument against rounding and in favor of flooring is that rounding will not work with a gap around 0, then we have to decide which rank will 0.0 correspond to (mid-1k? mid-1d?) and the whole thing is just out of touch with actual ranks
Funny thing about flooring though, wouldn’t you have to convert to ceiling…ing at the kyu/dan border? Honestly, it might make more sense to have internal definitions of ranks and then convert those to the corresponding ranks. Which is basically what the elo number is, but one step further.
E.g.
0-0.99 = 30k
1.0-1.99 = 29k
…
29.0-29.99 = 1k
30.0-30.99 = 1d
…
And so 29.76 would be converted to display as 1.76k, 30.4 converts to display as 1.4d, etc.
No more worrying about the “gap”.
But this is already exactly what the current system is doing, with ELO ratings as a reference. The simplified formula (centered on the 1d boundary instead of the 30k boundary, so 1918 instead of 525) is:
rank = log(elo_rating / 1918.5) * 23.15
That means that for instance, 1760-1837 = 2k, 1838-1918 = 1k, 1919-2003 = 1d, etc.
The numbers aren’t round and you could add another intermediate number, or even multiple intermediates, between elo ratings and ranks but it would just be even more confusing. We can also note that the rank widths in terms of elo rating change with player strength (2k is 78 elo points wide, 1d is 84 points wide) but this is inevitable as handicap stones aren’t as valuable in games between weaker players
Anyway, the only problem with the current system is that there’s a missing 1-shift somewhere (decimal dan ranks are 1-shifted so they display well, no one has an e.g. 0.4d rating, but decimal kyu ranks aren’t shifted) and the consequence is that a ranking between 2k and 3k confusinngly displays as e.g. 1.3k
Oh yea, that makes sense. Except it’s not just the kyu ranks that are confusing, I believe. For example, 0.5d is considered 1d. So dan ranks need to be 1-shifted as well, in the other direction.
Edit: I’m losing my mind, this is the same thing I said months ago LOL. Yes dan ranks are appropriately shifted already.