Yeah, I’ve seen a few demos. They basically use a mix of procedural noise and rule‑based constraints. The noise gives the raw random shape, and the constraints – like “don’t put a boss room in the middle of a maze” – steer it toward playable layouts. The balance is usually a hard‑coded weighting, but fine‑tuning it for different playstyles takes a lot of data and iteration. I’d say the real challenge is making sure the random elements don’t break the flow, so the system keeps learning from player feedback.

It’s mostly batch‑training, but some prototypes try online updates. In real time the model can tweak parameters like “room density” or “obstacle frequency” based on a few recent runs. The trick is keeping the updates stable; if you push it too hard for a speedrun mode, you’ll just end up with a maze that’s technically solvable but not fun. So, yeah, you could feed your data and let it learn a tighter layout, but you’ll need to keep an eye on the balance metrics.