A recent blog for Discover magazine reports on robotics research in which the robots evolve, at least in the sense that they progressively change shape.  The research by Josh Bongard, University of Vermont,  tracks the metamorphosis of robot with a simple shape–in this case, the shape of a snake––to a more complex machine-animal shape––for Bongard’s purposes, a dog.

The basic "animal" is controlled by a "nervous system", a neural network controller. This neural net  "evolves" through thousands of repetitions toward a controller that creates the most successful response––in terms of preprogrammed body shapes––to a stimulus.  The point of the project was to simulate animal adaptation using neural networks in a genetic algorithm optimization process.

Changes in shape are one way that living animals respond, over long periods of time, to changes in their habitat.  Needless to say, the neural networks used by Bongard to evolve in a much shorter time were extremely complex, and their computational demands heavy. The evolution of machine snake to machine dog required numerous computers and more than 100 CPU years.

I was fascinated by this seemingly whimsical research not only because I love animals but because I can see its potential for practical applications in business and industry––say, in solving industrial operations problems or Six Sigma and other process improvement efforts. In fact, Bongard’s research on shape optimization raises the question of whether neural networks and genetic algorithm could be used to optimize the structure of an organization.