I thought I could make a graph that shows how much time/game it takes to gain one rank. It would’t account for difference in effort, but still a fun little thing to do. However, ogs ranks are really spastic. Often times players gain/lose several ranks at really short time. So if someone stayed at 7k for month, and then got to 6k and 5k same day, then on the paper it would be 7k->6k takes month and 6k->5k takes almost nothing, which is counter-intuitive. I still made a graph, but I have no idea if it even makes sense. Maybe someone can think of something better.
BTW the results look intuitively right. We have this phrase “asymtotic DDK” which describes how most DDK’s graphs look. That is reflected in the increasing time to rank as you approach SDK.
Looks less spastic but still players gain several ranks at a time. Maybe we should make Fourier transform out of it, kinda looks like sinusoid. Example:
I would pick the axes “rank” and “percentage of games played at that rank”. I suppose ff you want a time-sensitive component, you can implement an “in the last x months” window.
If you want to get really fancy you could make it an xyz-plot (rank, % of games, unit of time); you’d literally get a “track” record. ;D If you want to counteract the bias resulting from people playing 1000 games one month and 0 the next, you can define each tick of time as x games played. Obviously that kinda requires a lot of datapoints.
Whatever the flaws of the graphs, the impression I get from them seems right in several aspects. The rise from newbie to strong DDK is a relatively smooth, steady curve reflecting fast improvement as elementary concepts are learned and mistakes overcome. The hump at about 12k to 7k nicely coincides with the transition from DDK to SDK, which I think many players would say matches their experience. Then, except for the outlier at 2k-1k, it levels off at a comparatively high interval, again matching what I have heard, that ranking up is much slower at the strong SDK and dan levels. In sum, I found this interesting.
Please link to example games of 20k winning against 11k (I suppose 9k can’t play ranked vs 20k ;D) in an even game, because short of sandbaggery, a case of “I have to go now and since you’re not resigning, I will”, unlucky timeout or “I think that’s enough, let’s review”, I don’t believe that’s possible. I hope you’re only counting ranked games.
I think IGS is much closer to the actual answer - if I were to play a 7k+ on even (9 stones difference), the system reports a 0.26% chance for the 7k+. That’s 1 out of 385 games. By the time 20k has played 400 games versus sdk, he probably won’t be 20k anymore. But 1 out of 20 actual games is still impossible. (The ddk line even suggests it’s closer to 1 out of 8)
Another perspective is to link this to the rank instability problem. If players are deviating within a 4~5 rank interval, then the 20k could be a 17k with a losing streak, and the 11k could be a 14k with a winning streak. Then it’s suddenly not so unlikely that the 20k wins sometimes from the 11k.
But it doesn’t work out like that because it would have to fluctuate for everyone. You might lose once in a blue moon to someone 4 ranks lower (I just scanned the last 400 games of one of the weakest low/mid-ddk i know and the most severe losses were 3 losses to people ranked 3,3,4 stones lower). If you lose 1% of your games to people 4 stones below your rank and someone else wins 1% of his games against someone 4 stones stronger, you’ll end up with a proportion of around 1:5000 (0.0002 or 0.02%) for 8 ranks (it depends a little on how you resolve ties since technically you could end up with a game that both are “pegged” to win, otherwise the figure would be 1:10000). Since I don’t have data for OGS I’ll use IGS again to estimate win/loss with 4 stone difference, which works out to 5%, which sounds somewhat more realistic to me. We still only get 0.25% winrate.
I strongly believe that the figures are inflated due to abovementioned reasons that have little to do with the actual game.
But I’m not talking about 9x9, I’m talking about 19x19. To include 9x9 is pointless because the difference is about quartered (I could win a 20 stone game versus 17k on 19x and lose a 6 stone game versus 21k on 9x). Can you reanalyse with purely 19x games?
And yes, to get a fairly clean estimate we have to exclude
Well, I was talking about all sizes. I even put it in the title
Though I understand you were arguing about something with flovo, maybe that’s why. I think it’s important to remember that only about half (actually less than half) games on OGS are played on 19x19. So it’s not like 9x9 is some edge case no one cares about, it’s a significant portion of OGS games. And rating system has to accommodate for that.
Yeah, it would be cool to have separate graphs for each size, it also would be cool to have separate graphs for each time control. But it’s not like 9x9 is pointless.
Should we exclude blitz too, since it’s also possible to have a lucky win there? I heard sometimes people resign in blitz by timing out, so we would be excluding those games if we want to exclude timeouts anyway.
Basically to redo the analysis with 19x19, I’ll have to redownload data. At least I don’t know any better solution. So it’s going to be a while. Number of moves is simply unavailable.
Well the ideal way would obviously be to analyze a single time setting (or each time setting separately unless that requires separate downloads), and probably most feasible to choose the one most people play (live?).
I know moves aren’t available, but it would be extremely helpful to have that available. API requests, anyone?
