Josh Bongard
Dept. of Computer Science, University of Vermont
November 9, 2009
12:50 - 1:45 pm
367 Votey Hall
Summary
One of the main criticisms of evolutionary algorithms is that much time is spent evaluating poor solutions. In this talk I will be presenting a method for greatly improving the computational efficiency of evolutionary algorithms by terminating solutions early when they cannot compete with the current set of offspring-producing solutions in the population. I will show how this method is independent of the fitness function used; it only stops those solutions guaranteed to become inferior to the current offspring-producing solutions, and realizes significant time savings across several evolutionary robotics tasks including legged locomotion and object manipulation. Finally, I will show that for many tasks, running time was reduced from polynomial to sub-linear time, and time savings increased with the number of training instances used to evaluate a solution as well as with task difficulty.
Speaker Bio
Assistant Professor Josh Bongard was awarded the prestigious 2007 Microsoft Research New Faculty Fellowship. His research centers on evolutionary robotics, evolutionary computation, and physical simulation. To find out more about Dr. Bongard's work, peruse his website, find out about his research on self-healing robots that aired on the Discovery Channel, or read his book, How the Body Shapes the Way We Think: A New View of Intelligence.
