Edit: The feedback from comments is that the names “teleological” and “ontogenetic” aren't good fits. I agree. There's an interpretation in which the terms work but there are also other ways to interpret them where it doesn't make sense. I had gotten the terms from the procedural content wiki, but in the future I'll avoid using these terms. I've changed the blog text to use “bottom up” and “top down”. Thanks for the feedback.
Long ago, I spent years working on the terrain generation for a game I worked on. I used earthquake faults, volcanoes, soil erosion, lava, river erosion, sediment deposition, water flow, weather simulation, vegetation (affected soil, which affected erosion), forest fires, floods, continental uplift, and other physical processes. This is the
teleological approach bottom up approach to making maps.
To make good maps, I'd change parameters and run it to see if the maps looked good. I'd go back and change parameters again and again. Ideally I'd write a search/evolutionary algorithm that explored the parameter space. To do that I need to write code to determine whether a map will look to my eye, and I never did figure that out. I knew it when I saw it.
The other approach to map making is
ontogenetic top down. Start with the structure you want to see, and make algorithms produce that directly and fill in the details. This is what I did with last year's simple map generation project. I had seen Perlin Noise, and thought, hm, that looks vaguely like a map — let's see how much work it takes to turn it into a map. Not much, it turns out!
For the past month I've been working on a new map generation project. I wanted island/continent maps with rivers, and Perlin Noise wasn't a good fit for that. I tried making random mountains come out of the ocean, and then using water flow algorithms to carve out rivers. I've been tweaking and testing, tweaking and testing, tweaking and testing, and finally realized that I'm back to using bottom up algorithms. It's incredibly addictive. It feels right: erosion, mountains, gravity, and all the other “realistic” things that form terrain.
The trouble is that I really don't need all of that. I'm working on 2D Flash games, and I don't need terrain; I need maps. It doesn't need to be completely realistic. It just has to be interesting. So I'm stepping back a bit and switching back to the top-down approach. I'm drawing some maps that I like, and then I'll find algorithms that generate maps like those, instead of modeling physical processes that produce realistic terrain. I think this will lead to better maps with less work.
Update: [2010-09-04] See my newer blog post to see the beginning of the new map generation project.