I mean, NP-complete problems are perfectly solvable computationally, and given the relatively small number of rooms per floor that architects are concerned with

I mean, NP-complete problems are perfectly solvable computationally, and given the relatively small number of rooms per floor that architects are concerned with, finding the optimal solution with a computer should still take much less time than it takes a human to come up with some local optimum in the design space. The real problem is actually encoding all the constraints that an architect is really working under, including such inherently difficult-to-quantify aspects as aesthetics, ergonomics, and familiarity! A solution that maximizes the available m^2 is useless if its human users find it too weird.