#procedural generation

For context:

Dwarf Fortress makes heavy use of procedural generation, Tarn has built a suite of algorithms designed to interact with each other in surprising ways to create unique stories for the people who play. He is good at this, in part because he is a trained mathematician able to build incredible complexity into those algorithms without breaking them. Dwarf Fortress is often cited as the strongest evidence we have that humans can create Art from an Algorithm.

Procedural Generation is pretty common in games, though, particularly when developers are trying to create large amounts scale or variation without completely filling up the user’s hard drive with assets they may never see. It can also serve as a jumping-off point; generate things algorithmically first, then go back over the result and add some flourishes so it feels more “lived-in”. No Man’s Sky is a good example of the former, and Bethesda makes heavy use of the latter. Both contain examples of how it can be employed poorly and effectively.

No Man’s Sky was very very “procgen” when it first released; a theoretically endless universe that basically looked the same everywhere you went. In a game where the main selling point was endless exploration and discovery this was a huge problem. They learned from their mistakes and tweaked their models and added adventure hooks until eventually we arrive at something most players find pretty compelling.

Bethesda games historically favor the “generate and glow-up” approach because it lets them build a large-scale sandbox quickly and frees up resources to focus on painting in more details so they can ship an almost-playable, sort-of complete game by launch day. You can see a lot of the seams of this approach in Fallout 3; it’s easy to get lost in the Capital Wasteland without the minimap because the environment tends toward the “procgen samey-ness” that you get when there’s not enough work put in to finish what the model spit out. There are even procedural artifacts encoded in the game: misaligned polygons where you can step into gaps in reality and fall through an endless void, assets clipping through each other jankily, locked off storylines that rely on an item variant that broke when the base model was tweaked. Yet it contains some fantastic environmental storytelling using fairly limited tools and manages to feel 100x more “lived-in” than, say, Starfield.

All that to say, before LLM psychosis gripped the tech world it was our old friend Procedural Generation who produced vast seas of slop for decades on end. “Bad Procgen” was instantly recognizable for its lack of detail and dull repetitiveness; you could be stuck slogging through it in shoddy MMOs and overambitious RPGs from the Commodore 64 to the PS4 and beyond. I see “generative ai” moving into that role more and more now with results that manage to seem sort of more convincing while somehow demonstrating less imagination than mathematics full of human error allowed run wild. There was a time when a Minecraft player could roam in one direction until the block generating code started to get confused and created an all-new biome they named The Far-Lands, full of floating islands and impossible cliffs. Even when it “hallucinates” a LLM tends to stick to the most plausible-sounding token predictions, I don’t picture LLM assisted generation managing such a feat of divergent wonder.

Stochastic Geomophical Transport for Terrain Erosion Simulation

One major way to model terrain has been through simulating erosion: how the rock weathers away is a big component of the vibe of the landscape on a geological scale. But there's a couple of components to that: both the erosion itself and also where that material goes afterwards. In short, this new simulation from Nicholas McDonald and Guillaume Cordonnier handles both mountains and rivers.

(Another major way, of course, is Perlin Noise and related approaches, which eschew teleological realism but gain other benefits.)

The idea here is momentum conversation: using a new particle-based algorithm (which can be combined with other geological processes, like tectonics and wind direction) it simulates geomorphological transport, which operates over geological time, taking advantage of the difference in timescales: over the course of geologic time, a river is basically instantaneous.

This makes it very flexible for mixing "a wide variety of phenomena" as they say: in the paper they describe the potential for things like dunes, coastal erosion, floods, rockfalls, varying erosion weights. I particularly like how effective it is at effects like braided rivers and river deltas, which are very common in nature but often overlooked on procedurally-generated maps.

On the other hand, if the erosion is fast (individual rockslides) or transport is slow (glaciers) that breaks the assumption and it won't be as accurate at modeling it.

I think the reason that I'm personally drawn toward this algorithm is because it has a history that is naturally embedded in it.

You don't necessarily need to replicate the exact phenomena that was involved in creating something to get a good result. Much of games and simulation is about picking the right abstraction to get the right feel, regardless of how you get there. It's often the better call, to get the right poetry instead of the exhaustively correct metric. But one benefit that you do get replicate the physical causative process to try to simulate the physical effects of water, wind, and time is that it comes with a built-in sense of history.

Me: