Design & train simple grammar

[mbforbes]

Check out simple grammar approach (e.g., L system paper in other repo) and plan how I'd learn the rules or probabilities.

Cool thing is can go into depth on whatever sub-problem seems the most interesting / fun. There are many sub-problems in procedural city generation.

Possible sub-problems to focus on (ones struck out may still need to be done, but aren't possible as an "interesting" ML approach):

• generate roads
• ~blocks~ (deterministic)
• ~blocks → parcels~ (this is a bad choice because I don't have parcel data; should either be latent, nonexistent, or deterministic)
• parcels / blocks → building bounding boxes (optional)
• building bounding boxes → 2D building shapes
• "" 3D building shapes
• 3D facades
• texturing facades

Something to keep in mind is that generating roads first, then parcels, etc.---what you might call a "pipeline" approach---was just done previously in handcrafted rule-based procedural generation, likely because it makes sense logically to computer people. It might not be the best tactic for an ML-based approach... and probably isn't even logically be how towns arise. It seems like functional areas based on geographic regions make the most sense. In fact, I wonder if this is a better project idea...

alt idea:

• 2d fuzzy map of broad regions (pick granularity and discretize into grid? hierarchical?)
• generate roads / blocks based on those, w/ blocks labeled w/ mixture
• blocks can then generate buildings according to mixture ... and maybe some learned grammar? likely not uniform.

NOTE: I should really read the "4d cities" paper---and possibly others---more closely to see if they basically gave designers these knobs. If the ideas of the knobs are good, then we can pick them as a representation to learn.

Ideas:

draft 1

• split grammar for block → parcels
• per parcel, can generate building. (a) simple rectangle, (b) generate sequence of rectangles, (c) learn some polygon rotating grammar, (d) generate waypoints directly. It seems like (b) and (c) would require some preprocessing, and (d) wouldn't, but (d) would also probably be a more difficult learning problem (or would it?).

maybe even 3d?

• https://osmbuildings.org has 3d data
• for most residential buildings, they're just rendering @ some generic height (~2-4 stories, probably). doing that would be as simple as writing a three.js exporter
• i wonder if OSM has building height data for any buildings?
• if not, osmbuildings.org does have a data export too, so could try to get 3d data from that (and just ignore textures for now)

random idea: webGL. recent inspiration: [http://demos.littleworkshop.fr/infinitown]

In back of mind:

• can reduce label space (e.g. just buildings instead of parks and land masses)
• could try to just populate spaces inside of road networks at first
• could reduce shape space (e.g. just to rectangles)
• could take into account geography nearby
• what explicit or latent variables to use / learn? (e.g., residential / commercial / industrial as higher-level category? might be more interestingly automatically induced)

[mbforbes]

Before this I need to extract blocks (#5).

I don't think parcels are available, but they may not matter. They could be faked by doing some kind of axis- (or block)-aligned voronoi diagram... or could just be fully latent? Ultimately what matters are where buildings are, so generating building bounding boxes would be enough.

[mbforbes]

Grammar approach ended up not being feasible for the time scale of this project. It turned out that doing a cGAN was, perhaps ironically, simpler. The above are all fun to keep thinking about, though. I'll leave this open for now.