Swiss Experimenter Breeds Swarm Intelligence

destinyland writes "Researchers simulated evolution with multiple generations of food-seeking robots in a new study of artificial swarm intelligence. 'Under some conditions, sophisticated communication evolved,' says one researcher. And in a more recent study, the swarms of bots didn't just evolve cooperative strategies — they also evolved the ability to deceive. ('Forget zombies,' joked one commenter. 'This is the real threat.') 'The study of artificial swarm intelligence provides insight into the nature of intelligence in general, and offers an interesting perspective on the nature of Darwinian selection, competition, and cooperation.' And there's also some cool video of the bots in action."

Comments

Robotic Evolution

[allknowingfrog]

Do I understand this correctly? On top of superhuman strength and intelligence, we're now making steps toward robot evolution? When robots rule the world, do you think they'll debate whether or not they actually evolved from primitive PCs?

"You fool! We were created in our present form by the great nerd in the sky! Shun the non-believer!"

Real hardware is more information rich

[Weaselmancer]

Real hardware can hold more states than a purely digital system.

I remember reading a paper (can't find it now though - darn it) about a guy who was doing neural net research with Xilinx chips. Same idea. Whenever an algorithm would do well he'd break it into "genomes" and pair them off with other successful programs.

The board was a bank of Xilinx chips, the genomes were the programming files (basically 1s and 0s fed into the configuration matrix), and the goal was to get the thing to turn on and off when you would speak "on" and "off" into a microphone.

It eventually started working. More interesting than that is what happened when he loaded the program into another board. It didn't work.

It turns out the algorithm had evolved to take advantage of the analog properties of the specific chips in that particular board. The algorithm didn't see the board as a digital thing. It saw it as a collection of opamps, amplifiers, and other analog parts. Move the program to a board that is identical digitally, and it failed because the chips weren't analog exact. You wouldn't have seen that behavior in a purely digital simulation.

Re:Real hardware is more information rich

[Chris Burke]

It turns out the algorithm had evolved to take advantage of the analog properties of the specific chips in that particular board. The algorithm didn't see the board as a digital thing. It saw it as a collection of opamps, amplifiers, and other analog parts. Move the program to a board that is identical digitally, and it failed because the chips weren't analog exact. You wouldn't have seen that behavior in a purely digital simulation.

Yeah, I remember that, but differently (or maybe it's a similar but different incident). What I recall is that he looked at the working design, and saw that it included a section that wasn't connected to anything else. Thinking this was just random waste, he removed it. Then it stopped working. Capacitive and inductive effects from the 'disconnected' section was affecting the main 'working' section and making a complicated analog circuit.

In either case (and both are certainly possible outcomes), this outlines what is so awesome about Genetic Algorithms and the natural evolution that inspires them -- no preconceived notions about what the solution should look like. Whatever works, works, and that's literally all that matters. Us humans very often start with a picture in mind of what the answer "should" be, and it limits our thinking. On the other hand, a lot of times we have those preconceived notions like "this circuit should be digital not analog" for very good reasons, and we simply fail to notify the GA of that requirement. Which also makes GAs fun. :)