It's official: robotics scientists are now just daring the things to kill us. An Aberdeen University have built a robot programmed to fulfill one objective no matter what, to evolve in order to do so, and to make use of extra materials when they're available. We're assuming the only reason they didn't call it "Terminator" is because they don't have enough to kill all the MGM lawyers. Yet.
You know how it took organics, hereafter referred to as "puny fleshlings", millions of years to learn to walk on two legs? The Incremental Evolutionary Algorithm (IEA) equipped robot learned in less than a day. When the research team added knees to the legs it re-learned, making use of what it had already learned about hips.
The robot uses a neural network to iterate solutions to problems. The evolutionary aspect is how it can just "decide" to add more neurons whenever the situation changes. The addition of new joints, limbs or even senses just creates a bigger, smarter and more-able-to-see-you robot. It can also lock itself into a specific configuration when it decides things are as good as they're going to get and unlock again as needed.
The system does have a weakness: if it tries to learn too fast, learning new behaviours with its whole brain instead of the newly created sections, it demands too much computing power and grinds to a halt. And that means we're safe forever, because it's not like computers ever increase in power.
Team leader Christopher MacLeod is now working on a system where the robot can not only adapt to new bodies, but learn to instruct its human slaves, sorry, "creators" in how many limbs and other components it feels it needs. We can only hope they know to say "no" when this list includes a Remington 1100 Autoloader, an Uzi 9 millimeter, and a phased plasma pulse laser in the forty watt range.
Posted by Luke McKinney.
TrackBack URL for this entry:
Listed below are links to weblogs that reference Unnatural Selection: Evolving, Improving, Implacable Robots: