Some of my professors said my Ph.D. project was too complicated to work, but tonight I finally got a tribe of learning animats to seriously kick the crap out of three non-learning tribes at the same time. Yay for me!
Red's the learner, in case you didn't guess. The white dots are resource points (evenly distributed on this map); the dark-colored squares are territory controlled by each of the four tribes (black is uncontrolled); the light-colored dots are individual animats (you can see three red animats near the center of this snapshot). There are five animats from each tribe, but some of them may be overlapping.
I'm curious how territory can be controlled by one of the tribes when there is no direct (or indirect through uncontrolled territory) connection between one of the animats and that territory square.
For instance, in the lower right area, there are blue squares, ostensibly controlled by a blue animat, but it appears the nearest such animat is in the center of the grid. It doesn't seem likely to me. Granted, the blue animats are non-learning and so may have an effectively random distribution of controlled territory.
Interesting results, never the less.
Ok - this looks completely fascinating. Is there someplace where you have a complete description of your project? I am very interested in seeing the ideas behind this as well as how it works in this kind of simulation.
Ooooh. Cool. Just made that pic the background screen on my PC at work.
AG: Discontiguous territories are seen in nature, although it's true that they aren't common. Note though that the learning tribe's territory is generally contiguous and concave.
KoF: You can see my proposal slides here.
Bill: Haha, ok. Looks like it'd be a bit too colorful for my tastes, and too distracting :)
Is there a way to "breed" animats in your system like in the Tierra project?
They don't reproduce, no; I'm not using evolutionary programming, I'm doing learning. Which is why it's interesting, in my opinion. Evolutionary approaches can reach results similar to what I got last night, but I haven't seen any distributed learning architectures for agent coordination.
I note that your red "learning" animats seem to be running together in a group. Is there a range beyond which they are unable to communicate with each other?
They don't actually communicate yet, but they can see each other at a range of about two squares. The flocking you see it entirely emergent, and quite interesting.