100% exactly. This is a very smart observation and is actually one of the main approaches I attempted first when tackling the creation of 'endless' mode. There are a few nuances and adjustments that would actually turn what you're suggesting into a somewhat viable approach, but I think in general you're on the right track with both the hypothesis and its limitations.
A lot of the early playtesting I did for 'endless' kept indicating, much to my initial dismay, that the continuity of the carryover white piece is pretty much essential. In hindsight, it kinda makes sense. Imagine playing Tetris or Temple Run but where every 30 sec you essentially start from scratch. I guess maybe that's why people are more likely to binge-watch a whole season of Breaking Bad back-to-back than a full season of Black Mirror?
If we didn't have that constraint for echo chess, we could've easily just saved the crowdsourced pre-labeled solvable levels into a big 'test bank' db, and randomly served one of them any time a player needs a new level to play. Think SAT/GRE questions that are all pre-curated beforehand but that get 'generated' at random for every participant.
Thankfully the ML classification approach I ended up going with completely sidesteps the carryover piece dilemma. It's like you said: you simply "generate [incl. carryover] and test for solvability". Plus it works, with 99%+ accuracy, and in real time. So I think we're good for a bit :)
A lot of the early playtesting I did for 'endless' kept indicating, much to my initial dismay, that the continuity of the carryover white piece is pretty much essential. In hindsight, it kinda makes sense. Imagine playing Tetris or Temple Run but where every 30 sec you essentially start from scratch. I guess maybe that's why people are more likely to binge-watch a whole season of Breaking Bad back-to-back than a full season of Black Mirror?
If we didn't have that constraint for echo chess, we could've easily just saved the crowdsourced pre-labeled solvable levels into a big 'test bank' db, and randomly served one of them any time a player needs a new level to play. Think SAT/GRE questions that are all pre-curated beforehand but that get 'generated' at random for every participant.
Thankfully the ML classification approach I ended up going with completely sidesteps the carryover piece dilemma. It's like you said: you simply "generate [incl. carryover] and test for solvability". Plus it works, with 99%+ accuracy, and in real time. So I think we're good for a bit :)