[hammock]

These types of exercises are useful not to find their solutions, but to reveal the shortcomings of the model used to create them - and our thus our own mental models!

Building design is older than humankind, and there's tens of millenia worth of reasons why they are the way they are. But those reasons aren't always clear. Essentially modern floor plan design is the output of a black-box machine (human) learning algorithm.

The flawed output here, and the comments, help us point out what some of the missing variables are: egress, airflow, construction cost, etc. Do enough of this and you can get a vastly improved model.

[elihu]

> The flawed output here, and the comments, help us point out what some of the missing variables are: egress, airflow, construction cost, etc.

...or maybe illustrate that some of the existing external constraints are stifling architecture as a useful art form?

Modern structures have to be fully designed before they are built, whereas most traditional structures were designed and built at the same time, and the design could evolve as needed to fit the environment. Now, we do everything with straight lines because it's easier to make plans and estimate materials and communicate with builders and file for permits and verify code compliance and so on, but you can do much more interesting and complicated designs if you don't have to communicate the design with a human at every step.

I think sometime in the not-too-distant future we'll have practical machines that can construct buildings designed by software to conform to the features and limitations of the building site and the desires of the future owners. If there's no need to communicate the design to a human other than "does this rendering look good?" then we don't need right angles everywhere just to make life easy for the draftsmen and carpenters.

This seems kind of related to something Christopher Alexander said about twenty years ago (which was linked from HN recently [1]), that current design and construction methods and business models have basically made good architecture well-nigh impossible to achieve on a wide scale. I don't know what he would think of using his architectural pattern language as a set of algorithm heuristics, but it's one possible way forward.

[1] https://news.ycombinator.com/item?id=17480178

[QML]

I disagree: the issue isn’t with the expressivity of the model, but with the objective problem and constraints set by the author. You can quickly adjust the algorithm to product floor plans resembling modern buildings by penalizing construction materials.

[googlemike]

How can building design be older than humankind?

[dsr_]

Many of the social insects -- ants and termites and bees, especially -- build "buildings" which have heterogenous rooms specialized for different purposes.

Genetic algorithms, indeed.

[hammock]

I imagine there were hominids building shelters before homo sapiens. And they probably took cues from animals.

[JeffreyKaine]

Weaver birds, beavers, bees, ants, termites, etc... Hell, we could even call shelled animals like snails and turtles mobile buildings :P

[vanderZwan]

> there's tens of millenia worth of reasons why they are the way they are.

Eh, not quite: all of these "reasons" are adaptations to a specific condition. Many of which only exist since recently (for example, reinforced concrete enforces certain adaptations).

Evolution is not working towards one static best solution. It is always a process of adaptation of an existing solution to the current context.

[hammock]

I don't disagree. However no one civilization would have been able to invent, for example, a skyscraper of reinforced concrete, in a vacuum, without everything that came before that.

[tomcam]

> Essentially modern floor plan design is the output of a black-box machine (human) learning algorithm.

Elegantly put. Thanks. I was struggling without success to make a similar observation.