Artificial Intelligence: A Learning Robot
by Sarah Edmunds
Ever since humans appeared on Earth, we have been evolving. We went from slouching, cave-dwelling hominids to upright, intelligent Homo sapiens. We developed tools, learned to farm, created machines to do work for us, and invented robots to operate those machines. Now, scientists have taken the realms of artificial intelligence a step further. Engineer Christopher MacLeod and his colleagues at Robert Gordon University in Aberdeen, UK, have created a robot that can learn and evolve by mimicking biological evolution.
Before this model, the smallest physical change of a robot - the addition of a limb or a sensor - would require complete software redesign. This new robot, however, contains complex software that is capable of adapting to change over time, much like the human brain. A set of interconnected processing nodes, much like the human neural network, can be trained to accomplish a desired function through trial-and-error.
For example, if the robot's purpose is to remain balanced, it will receive inputs from sensors when it begins to tip over and will adjust its limbs accordingly. The input signals telling the robot to move its limbs each have a certain amount of importance. Specific combinations of these weighted signals result in a command to the node, telling it to drive a motor to move a limb, for example. If the action is successful, the combination is stored within the robot. If the combination fails and the robot tips over, it will make adjustments and try a different combination next time.
But the robot's interconnected processing nodes can do much more than solve simple problems like balancing. MacLeod's team of engineers developed an incremental evolutionary algorithm, or IEA, that is capable of adding new parts to its robot brain over time, spinning artificial intelligence into a new dimension of possibilities.
At first, the IEA robot was simple in construction. It had two legs and one primary purpose to its life: to travel as far as it possibly could within one thousand seconds. The robot fell over at first - much like a baby learning to walk - but got back up and tried again. Eventually, it stopped falling over and maintained stable progress forward, thereby completing its primary command. At this point, a conventional robot would have completed its purpose. Any addition of limbs or joints would have required the robot to start the learning process completely anew. The IEA robot, however, memorizes the first neural network that allowed it to walk, and builds from there. With the addition of knee-like joints, for example, it uses previous knowledge of how to control the top-half of its legs to gain control over the newer legs below the knees. Once it overcomes this second challenge, it memorizes the new successful neural combination so the process can be repeated. The robot's brain learns in small increments, mimicking the way the human brain evolves in layers.
This robot's ability to learn - or rather, the software's ability to slowly improve itself - seems like a horrifying cliché in a science-fiction movie, when the mad scientist realizes his creation is "alive!" and has, without permission, evolved into an intelligent being. But if a machine can evolve, what is stopping it from thinking for itself? Maybe that fear isn't so far-fetched after all. The invention of this new robot says one of two things about human intelligence. Either the process of human evolution is so basic, so completely transparent and understandable, that a simple robot can learn in a matter of hours what living organisms learned over the course of millions of years. Or, conversely, the human race is, in fact, so ingenious that we have mastered the inner workings of our brains. Our expertise about our own species is so precise that we can recreate the process of self-evolution within machines.
Or maybe it's neither. This advancement in technology is certainly profound, but perhaps it is the merely the logical next step in our own evolution. Every step we take in science is arguably just the next layer in fulfilling our "primary command" as humans. Each advancement just feeds our natural inclination to broaden our understanding of ourselves.
