So there's plenty of information out there that can take you from a beginner to a good player. However, the skill gap between good players and the best players is enormous. And paradoxically it's rather subtle and challenging to describe. Stuff just sort of happens so fast that even the best players aren't particularly good at articulating why they are the best. I would say that the whole upper range of the skill curve is largely mysterious, and that this is due to the fact that a majority of the progress boils down to "magically better intuition". So I've been thinking that we could actually do something about this. The skill curve really does simply boil down to mastering numerous quantifiable things. All we're missing is the data to do it. That's where this project comes in. In my opinion, an attractive target game to approach is TGM1, for several reasons: Thanks to The Masters, there is a wealth of superplay videos out there There is a good range of players, each with reasonably well defined relative skill levels Because of the competition's nature, we see representative samples that don't trade risk for reward the way a world record might The plan is to create a system where a video could be turned into a replay file. The replay file would be compatible with eg. MAME, however it would instead be played in a basic non-graphical engine that could tear through the replay much faster than realtime. This engine would also let you program any "hooks" you want so that you can log variables of interest. The whole idea is that you could add new variables to try to answer new questions that weren't even thought of initially, and this would be no big deal. The slow video -> replay conversion is only done once. Afterwards, the engine efficiently reprocesses all replays in the database to calculate new stats super fast. I'm not talking mere piece counters or APM/TPS here, hopefully some of you have the imagination to see just how great the possibilities are. The engine would of course process my and your shitty replays as well, and articulate all 1000 dimensions along which we suck. Call me a masochist but I get tingles just thinking about it. Research has shown that practice isn't nearly as beneficial as focused practice, and this would provide the mother of all auto focuses. You could also do stuff like take your lossy video recording, and remaster it into a razor sharp emulated video. Would be cool to remaster the world record games etc. The engine alone would be ambitious enough, but this requires so many things before it even comes to that: Reverse engineering the randomizer (done) Code to find the randomizer seed given a sequence Some machine learning thing to translate video into game state information Code to impute the replay sequence that lines up with the game states A frame perfect engine to run the replay files A bunch of hooks to log interesting stats Then the fun finally starts As a project, I have a decent idea how to move forward with this, and the bits that are less clear synergize well with stuff I want to learn for work anyways so I could see myself chipping away at it in my spare time. Does anyone else have interest in collaborating / have the required skills?