I’ve started to collect some data on my own play and that of my opponents in order to look for patterns in my mistakes and how those patterns stack up against the competition. Things like:
Missed nets / ladders
Direction of play mistakes
Misjudging sente
Misreading cuts
and so forth. Some of this I can automate, but a lot of it is my own manual annotation, which is a bit tedious. Does anyone know of a tool out there that can tally up this sort of thing for you, and compare your performance to players of similar ranks?
Also: I don’t have a ton of data yet, but early indicators suggest that there is a MASSIVE increase in quality of play from 1k to 1d, in the sense that 1d’s make much higher quality mistakes than 1k’s. My rank is 1d, so most of my opponents range from about 4k to 3d, and I would say that the difference between 1d and 1k is bigger than the difference between 1k and 4k, and bigger than the difference between 1d and 3d. Maybe people covet the dan rank and play a lot more cautiously when they get it, or something?
Really? I’m currently 1k, was 1d a few months ago, and I don’t feel I play much better when my rank is 1d than when it’s 1k. And I know that dan players sometimes make big mistakes too.
Are you using AI in some way to identify mistakes, or is it by your own judgement?
I think that even defining what is a mistake is not easy, even more so when talking about higher / lower level mistakes etc. To do a rigorous analysis, we need to first establish what exactly we understand by that.
However, to get a better understanding of the difference between ranks, I believe it would be sufficient to analyse the ranking system that produces these ranks.
What is funny is that in every ranking system (KGS, OGS, AGA, EGF,…) people think there is a big qualitative jump between 1k and 1d… However the threshold 1k-1d is different on each server/go association, and it also changes over time.
Well, of course the stronger you get, the harder it is to gain one rank, so steps become higher and higher. What I’m saying is that steps don’t brutally become higher between 1k and 1d. I’m 2k on the FFG/EGF scale, and feel that 1d players don’t play very differently, just a little better in most aspects (better joseki knowledge, more precise reading for instance). On the other hand, high dan players seem to play a different game, they see many sequences that would never come to my mind.
Are you using AI in some way to identify mistakes, or is it by your own judgement?
It’s a combination of both - I’m using the AI to review the game, but my own judgement to mark and categorize mistakes. And obviously my judgement here is not 100% fool proof, but there is a clear difference between, say, missing that some stones can be captured in a net and misjudging subtleties involving sente in the early middle game.
However, to get a better understanding of the difference between ranks, I believe it would be sufficient to analyse the ranking system that produces these ranks.
The ranking system that OGS uses could be applied to any 2 player game with wins and losses. It can’t tell you anything about what skills you need to improve in order to compete at the next rank.
People say the same about 9/10k.
There are strong 10k’s that will swear black and blue they aren’t strong enough to be SDK.
It’s the same with 1k/1d.
I haven’t collected any data about 9k/10k, but in the 4k-3d range where I do have data, the differences between ranks are not in fact the same. I dunno what to tell you here - I performed an experiment and I’m reporting the results.
I haven’t done any analysis on other servers, so I can’t confirm or deny. But one possibility is that there is a big psychological barrier between being a kyu player and a dan player, so a 1d player puts more effort and attention into their games than the same player would if the ranking system were shifted so that they became 1k. No idea if it’s true, just a theory.
Well, so far all you really presented is unsubstantiated claims. Do you have any data from these experiments you can share so that the community can compare your experiment methodology, input data, and results with your claims? There are many clever people in these forums who have done many different forms of analysis in the past comparing user ranks, I am sure many could provide good insight into your study if you provided more details.
Well, so far all you really presented is unsubstantiated claims. Do you have any data from these experiments you can share so that the community can compare your experiment methodology, input data, and results with your claims?
So far all I have actually done is manually marked up a couple hundred games to look for patterns that I can use to improve my play - I wasn’t trying to announce a major scientific breakthrough. It would take some effort to organize my annotation into a form that could be easily shared with others, and that’s not really my goal right now. If you would prefer to assume that I’m lying or incompetent pending further evidence, fair enough!
I’m not sure about tallying similar mistakes, but I believe [AI Sensei] lets you analyse your game based on a given rank. This may be helpful?
I use an offline version of KataGo rather than AI sensei, but the problem is the same either way - it can tell you which moves are mistakes and their objective point value, but they can’t categorize mistakes in human-comprehensible terms (“missed a net”) nor can they measure the level of the mistake on a practical human scale (“4k’s often make that kind of mistake; 2d’s rarely do”)
Understanding the skills you need to improve is not necessary to measure winning probability.
I’m confused - this seems… backwards? My goal is to improve my ability to play Go, and to achieve this I have analyzed patterns in the play of other players near my strength. If I did an analysis of win probabilities then I expect that I would find that 1d’s are more likely than not to beat 1k’s, 2d’s are more likely than not to beat 1d’s, and so forth. This would not be particularly surprising, and more importantly I’m not sure how I would use this information to get stronger.
Given the title of this thread and your claim that the difference between 1d and 1k is bigger than the difference between 1k and 4k, I assumed that this thread is about the difference between ranks. However if that is not the case, then please ignore my posts.
Given the title of this thread and your claim that the difference between 1d and 1k is bigger than the difference between 1k and 4k, I assumed that this thread is about the difference between ranks. However if that is not the case, then please ignore my posts.
I tried to make my intent clear with the first sentence of my post: “I’ve started to collect some data on my own play and that of my opponents in order to look for patterns in my mistakes and how those patterns stack up against the competition.”
But I guess I wasn’t clear enough - apologies for the confusion.
To check that I understand what your initial post was about: you analyzed your games against opponents of various levels, and you categorized mistakes into several types. Some mistakes are “big” mistakes, others are “small” mistakes. You made the observation that the number of “big” mistakes drops a lot from 1k to 1d. Did I understand correctly? If so, can you give a few more examples of big mistakes and of small mistakes? And approximately how many big mistakes per game did you find in 1k games and in 1d games?
its also possible to measure difference in play style between ranks without AI.
just collect 100 games where 9k played as black
and 100 games where 1d played as black
then look which coordinate was the most popular on move 5 for example.
it would be interesting if coordinate is different between ranks
it also possible to detect on which move number the first 5th line move was played, then you would know how fast you should go into the center…