So as I'm trying to get better (and succeeding ever so slowly) I've been thinking about what kinds of training would help. I could imagine a Tetris version that gives you real time feedback about whether or not the placement of your last piece was good. "Good" could be modeled simply to begin with, maybe something like 'does your placement increase or decrease the number of possible next pieces that could be stacked smoothly' and grow more complex to incorporate things like knowing what the next piece is, spins, and more abstract goals. I know there's a lot of special cases and situation-specific goals, but the aim would be for a helpful model, not a perfect one. Obviously it would run at low G to let you process the feedback, but the goal would be to use it to get into good habits more than to train for speed. Has anyone made something like this yet?
There's nothing like that for stacking as far as I know. I'd be impressed if there was - stacking is somewhat subjective and heavily dependent on what you're ultimate goal is (stacking fast vs stacking for Tetrises vs t-spins etc) and also what rotation system you're using. It's a complex problem that people have put a reasonable amount of time towards in building Tetris AI, but I've not seen anything that takes it further than that. To be honest, I'm not sure how helpful it would be as a training tool - I feel like it'd be something where you could waste a lot of time trying to understand why a semi-arbitrary AI said your placement was bad rather than just playing the game more and getting better through practice. High level Tetris is almost entirely reliant on instinct and experience, neither of which are helped by concentrating on such a high level of detail. The principles of good stacking are already pretty well documented in various guides for new players, and I'd say knowing and understanding the principles are all you should have to do. Looking at individual placements on a case-by-case basis can be informative, but the combinations of stack shapes and pieces (and next queues) is so high that I don't think it would help you improve much as a player unless you can memorise thousands of things having only seen them once or twice, and then recall them instantly the next time you see them... I'd see it as similar to learning mahjong, poker, chess, etc. The good players *know* (or, more accurately, calculate very quickly) what to do in any given situation, but they've done that through thousands of hours of experience and by gaining a deep understanding of the game and how their decisions affect the future of it. Not by individually learning each scenario and which choice would be the best one. On the other hand for improving things like speed and finess, afaik there are versions of Nullpomino which have some pretty helpful tools to gauge how efficient you're being.
Yeah, I looked up some of the AI work. It's interesting (there's a cool GA video that got me thinking along those lines), but in most cases seems to punt on the more complex goals of the game and just reduces it to clearing lines for survival. I think there might be a middle ground between that and the instinctual play that you're talking about where considering various goals might be useful at the level of individual placements; kind of a moneyball or DVOA statistical take on the game. I wouldn't want it to be so slow that you, say, stop the game between each piece, but letting the player know if they missed a better placement without disrupting the flow of the game seems like it could be useful. (Probably it would be combined with replay so you could see what the game saw that you missed.) If nothing else, it's an interesting algorithm and interface problem. I'll have to check out the Nullpomino stuff. Is it built into the main game somewhere and I just missed it, or are there separate forks?
The Nullpomino stuff as I know it is here: http://harddrop.com/forums/index.php?showtopic=4162 Though that might be outdated. I think the better up-to-date version is incorporated into the Nullpo League fork: http://www.celer.be/. At the very least one of the screenshots has a "Finesse" value next to the playing field. And I think a moneyball type stat would be nice after games, but I still don't think it's actually all that useful as a training tool. Maybe you could see your own improvement in stacking over time, but I'm not sure in terms of looking at each placement it would ever grant you much improvement. Moneyball works when you've got time to sit down and analyse the data before making a decision, in high-speed Tetris you don't have that.
There's some nice stats available through that. Example: http://novomancy.org/tetris/2015_02_10_20_49_59.html (Comparing my numbers to others that are floating around is only mildly depressing.) I'll have to play with it to open up other play modes since they have it locked into line race by default.
Playing with no hold, perfect finesse and only stacking tetris's should be enough practice for stacking well @ 0G
