Polyworld Movies

If you think you might wish to view any of these more than once, please consider downloading them to your machine (option-click on Mac, right-click on Windows), rather than streaming them multiple times. Thanks. (They need to download fully to your browser cache before they start to play, so if you try to play them directly from the web site it will take just as long as downloading them anyway—every time!)

Modern Movies

These videos are from the modern incarnation of Polyworld, recorded digitally to hard disk in my lossless but highly compressed (180:1 or so) native format, then converted to QuickTime movies using Apple's Animation codec at the highest quality setting. They are much higher quality than the historical movies below, but only show the main oblique view of the world, and I so far have only converted a few samples.

First up are the built-in primitive behaviors in a toy world constructed solely for the purpose of obtaining these sample movies:

• Eating (1.4 MB) – Watch the initially orange, top-most agent as it moves over a block of food and consumes it. The agent is also modulating some of its other behaviors, including the brightness of the polygons at its front.
• Killing & Eating (1 MB) – Observe the two orange agents at the left. The upper agent of the pair catches up with the lower agent, kills it—at which point it turns into a block of food—and then consumes some (but not all) of the new block of food.
• Mating (3.5 MB) – Pay attention to the two agents on the right. When they come in contact, they produce an offspring that heads off towards the center of the world. Note how the child behaves more like the salmon-colored parent (in terms of color and movement), but has a body plan (size and aspect ratio) more like the other, blue-purple parent, due to a mixing and matching of genes.
• Lighting (0.7 MB) – Notice the left-most, slightly bluer agent. As it turns towards the center it also dims the brightness of the polygons on its front. This is achieved by the agent reducing the activation of its "light neuron".

And now a "test"… Watch the agent that is initially purple and on the left in this movie (4.4 MB) and see which three primitive behaviors you can spot, in what order.

Here are some excerpts of three different simulation runs, showing the evolutionary progression over time of different behaviors:

• Dynamic Food (excerpts) (33.9 MB) – This was one of a few experiments testing to see if the agents would evolve to track moving food. All food appears in one quadrant of the world for a short time, then moves to the next quadrant, then the next, the next, and then we repeat. Initially the agents pretty much ignore the food (which, of course, isn't too evolutionary useful) but over time they evolve behaviors that bring them fairly rapidly to the food patch each time it moves.
• Complexity w/ Barriers (excerpts) (23.4 MB) – This is a low resolution (640x480) recording of one of the runs examining evolutionary trends in complexity, as reported in this ALife XI paper. 80% of the food is distributed in a patch that occupies 40% of the depth of the world, on the end of the world that the barriers don't quite block off; the remaining 20% of the food is distributed in a patch that is 10% of the depth of world, deep in the end of the world fully broken into thirds by the barriers. (If that description of the world isn't clear, just look at the movie; it should be obvious what is going on.) Other experiments (not shown here) demonstrated that without evolution the seed population could not thrive in this world, with the agents always suffering a non-recoverable population crash. But allowed to evolve, as shown here, the agents fairly quickly evolve simply strategies that keep them near the primary source of food (the large patch), and over time also evolve to exploit the secondary, smaller patch. As evolution proceeds in the world the agents' neural complexity increases. (Complexity is measured by Sporns's information-theoretic complexity measure, C, a simplified approximation to Tononi, Sporns, and Edelman's C_N; for more information see this ALife X paper, the previously linked ALife XI paper, and their references.)
• Complexity as Fitness Function w/ Dynamic Food and Growing Barriers (excerpts) (52.7 MB) – This is one of a small number of simulations in which I used neural complexity as a fitness function, and used Polyworld in a steady-state GA mode. It happens to also be one of the dynamic-food setups, with a different pattern of food growth from the simulation above, and used a slowly growing barrier that gradually isolates the different dynamic food patch locations. Natural selection in Polyworld has so far not done a great job with this particular setup. But what is interesting here is that even though neural complexity grows dramatically (roughly three times as large as it ever grew in any of the above “Complexity w/ Barriers” runs), that complexity does not translate into evolutionarily useful behavior. The agents evolve different behaviors over time, but never do pay much attention to the food. This demonstrates that not all forms of neural complexity are evolutionarily useful, and suggests that studies of the complexity of network dynamics may need to take the nature and source of that complexity into account.

Here are a couple full-length movies of complete simulation runs:

• Dynamic Food (463.5 MB) – Full length source of the Dynamic Food (excerpts) movie above.
• Complexity w/ Barriers (197.3 MB) – Full length source of the Complexity w/ Barriers (excerpts) movie above.

Historical Movies

All of these videos are from the original incarnation of Polyworld, which ran only on a Silicon Graphics workstation. The videos were recorded in a variety of primitive fashions, and are generally of fairly low, but viewable quality. I have (relatively) recently encoded them as QuickTime movies with the modern H.264 codec (which means you may have to update your copy of QuickTime to at least 7.x), so they look about as good as they can given the quality of the original video recordings.

Modern experiments with Polyworld are digitally recorded in a lossless but highly compressed (180:1 or so) format I designed for that purpose and require the PwMoviePlayer app to be viewed (unless I've manually converted them to QuickTime movies, like the short ones above). On the one hand they are drastically better quality than these old video recordings, but on the other they currently only capture the main oblique view of the world. Maybe one day I'll put some newer videos of full-length simulations up, but here are the older, more historical Polyworld videos.

Built-in primitive behaviors in a toy world constructed solely for the purpose of obtaining these sample movies:

• Eating (0.2 MB)
• Killing & Eating (0.3 MB)
• Mating (0.4 MB)
• Lighting (1.1 MB)

A few specific evolved behaviors:

• Visual Response (3.6 MB) – very early example of agent using its vision input to modify its behavior
• Fleeing Attack (2.8 MB) – very early, very poor quality sequence showing an agent first ignoring an agent that doesn't attack it, then running away from an agent that does

Generic simulator sequences:

• Polyworld sample (6.6 MB)
• Brain dynamics (2.8 MB)

Early primitive “species”:

• Joggers (1.6 MB) – The very first successful population. Uncomplicated world with no barriers, wrap-around borders, and plenty of food and agents/potential mates, so evolution, as usual, discovered the simplest possible solution: always run straight ahead, always want to eat, and always want to mate. It was enough.
• Indolent Cannibals (8.3 MB) – A typical "cannibal" population. Really they were more like lazy, umm, maters. These were the only kinds of populations I got for a while, to the point that it began to worry me. Turned out I had unwittingly provided a particularly easy way for them to find both mates and energy… By (at the time) not requiring the parents to contribute any energy to their offspring, and gifting each newborn with a good supply of energy, children became an excellent energy source. So these agents lived with each other, mated with each other, killed each other, and ate each other when they died. Why leave home? After introducing an energy budget, so parents had to give up a genetically determined fraction of their energy to their offspring, I still sometimes see clusters and swarms (moving clusters), but never again the static, unchanging "cannibal" populations.
• Edge Runners (3.8 MB) – An example of another early species that "took over a world" by behaviorally isolating itself from the rest of the population. (I say "take over the world" because they were fecund enough that they fairly rapidly reached the imposed maximum population limit, thus preventing agents exhibiting any other behaviors from reproducing.) The video starts before the "edge runners" have completely taken over, but the small agents that just run around and around the edge of the world soon take over. There's a time jump to a point at which an interesting mutation in their behavior has taken place: They've evolved to run most of their lives, but stop late in life, so they can whip out three or four offspring with other edge runners passing by.
• Dervishes (8.1 MB) – My first (and so far only) "Braitenberg table-top" world, where the edges are open and dangerous… If an agent runs past the edge of the world, it dies instantly and is removed from the simulation. The agents evolve to simply turn in modest circles, thereby avoiding the dangerous edges of the world while still moving around enough to find food and mates. After a while the video jumps to some really awful quality video (recorded by pointing a video camera at the computer screen and recording its signal on a time-lapse security VCR). The poorer quality video is included because it shows a sort of continuous tit-for-tat behavior in the different populations on either side of the long barriers. Agents in a given domain tend to all express their fighting behavior or not, with the current strategy being invaded by agents mixing in from the barrier gap, at the edge of the world, leading to the adjacent domain. You could see waves of color sweep through the populations as their dominant strategies changed. If I remember right, this one ran for something like 1,000 generations. I should have probably reserved the moniker, "dervishes", for some of the agents in the next simulation, however…
• Foraging & Swarming (12.1 MB) – The last decent quality video from a run of the original incarnation of Polyworld, and by far the most interesting. By this stage, the individual agent behaviors have ceased to be one-dimensional; i.e., I can no longer sum up the total life behavior of all agents in the simulation with a single word or phrase. Here we see agents foraging—orbiting food while they eat it, despite there being nothing built in to attract them to food. And a swarm of small agents that stay together, which is good for finding mates, even as the swarm drifts along, which is good for finding food. Look around the world, at the range of behaviors. The population is sustaining itself with its mating behaviors, staying right at the maximum population limit, yet it is no longer obvious what the full range of behaviors is. The next step, then, is to develop the statistical and information theoretical tools to quantify these behaviors and neural architecture and dynamics of the Polyworld agents, which is what my modern research agenda is all about.