Hi there, while thinking about how I can improve my games and analysing my biggest weaknesses, I was wondering about statistics in my personal games. Not in terms of how many times I won or lost, but more in the sense of:
are there correlations between number of stones captured/lost and win rate, i.e., hinting at biggest problems may be too many weak groups, but overall tactics are fine?
is there a strong correlation between my average move time or game duration and the win/loss rate, i.e., hinting at my biggest problem being under pressure in byuyomi or not taking enough time beforehand.
correlation between number of “owned” corners by the end and the win/loss rate could point at practicing a more territorial style more than an “okay, I will just make an epic center” style.
I think it would be awesome to have a button to download a bunch of such fields, that could potentially be easily collected, as a csv. Or even, as a next step maybe, have such evaluations integrated into the website?
Thinking further: if one is able to quantify certain playing style aspects “on the fly”, one could even show a player his/her evolution of playing styles over time. (without having to value one over the other, just out of interest)
Thanks a lot for reading this far, and hopefully, I could spark some interest with this idea
What you’re asking for seems difficult to automate because those concepts are beyond the capabilities of current go AI. AI don’t neccessarily apply those human concepts while doing their computational thing. They might, but we really don’t know. AI are still very much a black box.
And those concepts are probably too fuzzy to capture in hand codes rules in an expert system. It’s not as if programmers never tried, but they never really succeeded at it with go.
To me it seems that you’re looking for a human coach/teacher/study group. As an alternative, you could ask for human game reviews in these forums or on GoKibitz.
I think a better use of the limited and precious OGS development resources would be a feature to download all your games as a big zip file (I and others have requested this on the past, search the forums, there is a monster zip of every OGS game ever but I don’t want to download 6GB just to get my 100KB, my desired use case is pattern search of my games) and you then do this analysis in a third party sgf database program like kombilo, or some custom scripts, which are written by someone other than anoek.
You could ask somebody who has downloaded those games to filter your games out for you.
You don’t really need AI for statistics like
At least for games that go to scoring there will be a well defined number of captured stones and a player that won, so you could count those up.
However how many stones were captured in a game that was resigned probably needs a score estimator to guess, Katago or something else - but even that might be debatable whether something that’s captured would’ve actually been captured comparing the level of the players to Katago.
Similarly no ai needed in principle.
Again in resigned games there might need to be some guesswork. If someone resigns when there’s only a 4-4 stone in one corner, should it be counted?
Lots of questions and choices to make it one wants some of these statistics.
I actually did this manually on my games many years ago. There was a noticeable difference in the average number of corners I controlled (which I defined as owner of the 2-2 point, scoring half if dame): IIRC in the games I won it was over 3, whereas ones I lost 2.5 or so (my overall winrate was ~90% so a smaller sample size). So still more than 2, I was a territorial player, but the lesson seemed to be if you wanted to beat me your best chance was to play territorially and try to make me build something in the centre because I was relatively weaker at that, or perhaps that gives more chance for a high stakes fight game versus a calm territory one where I’ll beat you in endgame if I haven’t before then.
This was all before AI and the direct 3-3 invasion.