Behavioral diversity has evolutionary benefits... in robots
A new study suggests that by encouraging behavioral diversity, we can produce a faster, more effective learning curve for robots attempting to complete designated tasks. The study results were published by Jean-Baptiste Mouret and Stephane Doncieux in the journal Evolutionary Computation, produced by the Massachusetts Institute of Technology.
Evolutionary Robotics is a field that aims to create robots able to learn and evolve independently, without anyone dictating their specific course of development. If you assume that robots (or algorithms) will one day take over the world, then evolutionary studies may be our best bet for understanding and controlling our own future. On the other hand, it may also be how we enable our own demotion in the man/machine hierarchy.
In either case, this most recent study shows how we can improve robots' evolutionary progress by focusing less on diversity in genotype design, and more on encouraging a range of behaviors. In other experiments where different genotypes have been tested for problem-solving fitness, researchers have found a tendency for robots to converge quickly on a "best-candidate" solution. That rapid convergence naturally limits exploration and the possibility of further growth. However, in the Mouret/Doncieux study, robots who were able to extend exploration when an array of behaviors was encouraged showed substantial evolutionary improvement "regardless of genotype or task."
From the study report:
When behavioral diversity is fostered, many tasks that were previously hard to solve become easy to tackle. This makes these techniques a valuable improvement to the state of the art in evolutionary robotics.
I've seen no specific evidence that we can apply these robotic findings to our own behavioral development, but in my mind, they sure do put "teaching to the test" in a whole new light. Perhaps we should be less focused on encouraging task completion in young students, than on encouraging a wide range of problem-solving approaches. It appears to work for the robots.
This post was originally published on Smartplanet.com