Ok, I really like writing these things. I think I'll try to do this weekly again.
Mad Reference: Josh Bongard. "Morphological change in machines accelerates the evolution of robust behavior." Proceedings of the National Academy of Science. January 25, 2011 vol. 108 no. 4 1234-1239. (abstract)
A lot of my random reads come from "cool" via Recommendations, a Google Reader list of things that other people have shared. As far as I can tell, this one mostly made that list because the title was cool, but the research is pretty neat.
Mad Background: Evolutionary algorithms are just about my very favoritest of things. Basically, computer scientists, inspired by the simplicity and elegance of biological evolution, have started using reproduction (of code fragments), mutation (of the specifics of the code), and selection (of the code fragments that are most successful at whatever task the programmers give them) to evolve programs. Such evolutionary algorithms often find solutions that the programmer might not have thought of, and do so faster than would occur if the programmer directly designed the solution. In other words, evolutionary algorithms flip the bird to intelligent design "theory" (which is only a theory in the colloquial sense of "guess;" there's no science there to actually make it a scientific theory).
Evolutionary algorithms are often used to design robots, both real and simulated (as shown in the video above).
Mad Observations: Real biological systems don't start out even crawling; they go through various body plans through their lifetime, and through the many lifetimes that take them from one species to another. For example, life has, multiple times, gone from a snakelike slithering form to legged walking form, both over many generations and within the lifetime of single organisms.
Mad Hypothesis: A robot that starts in a snakelike ("anguilliform") body plan before progressing to a four-legged walker can figure out how to walk faster and better than a robot that starts out up on four legs.
Mad Experiment: This experiment was mostly virtual. The experimenter set up digital "robots," some of which started out with legs, and some of which started out down on their belly. He then allowed them to search for a control mechanism for their four limbs that would get them from one corner of the simulation to the other. He also ran trials in which the belly-walkers gradually stood up on their legs (by slowly increasing the angle between the legs and the body), and then ran trials where robots pre-loaded with the slithering programs started out up on their legs. He even built Lego robots similar to his digital creatures, which just makes it all cooler.
They All Laughed, But: It worked. Robots that started out up on their legs took several hours to figure out how to walk, while the slitherers (and their progeny) figured it out in seconds. Even the Lego bots, which were slowly changed from slitherers to walkers, were able to figure out how to walk in the real world really fast, much faster than their always-walking cousins.
Mad Engineering Applications: When you're building your robot death army, don't start them out as bipedal killing machines. Sure, give them arms and a desire to rip things to shreds, but start out forcing them to use those arms to slither, then walk, and then finally they can stand erect and use those arms for their intended purpose.
This would probably also be useful for less destruction-focused robots, but that wouldn't be nearly as Mad.
No comments:
Post a Comment