Efficient and Responsive Behavior Design by Selective Evolutionary Generation
This talk introduces Selective Evolutionary Generation Systems (SEGS), which are a class of self-reproducing systems that are useful for tackling a generalization of the standard global optimization problem known as behavior design. Instead of an optimizer, a SEGS produces, on a search space, a probability density function called the behavior. The generalization depends on a parameter referred to as the level of selectivity, which restores traditional optimization when the parameter equals infinity. The motivation for this generalization is twofold: traditional off-line global optimization is unresponsive to perturbations of the objective function, and on-line optimization methods that are more responsive can be computationally expensive. This talk explains how a SEGS produces rational behavior via a Markov Chain Monte Carlo method to obtain both efficient optimizer search as well as responsiveness. The canonical genetic algorithm with fitness proportional selection and the (1+1) evolutionary strategy are particular cases of a SEGS. The SEGS method is illustrated through the evolution of flapping wing gaits in a way that is responsive to changes in flight conditions.
Amor Menezes is a Research Fellow at the University of Michigan, where he recently completed his Ph.D. degree as an NSERC Fellow and Michigan Teaching Fellow in the department of Aerospace Engineering. He received a Master of Science in Engineering degree from the University of Michigan as a Milo E. Oliphant Fellow in 2006, also in Aerospace Engineering, and graduated from the University of Waterloo with a Bachelor of Applied Science degree in Mechanical Engineering with Distinction and Dean's Honors in 2005. Dr. Menezes’s research interests include the fields of self-reproducing systems, stochastic optimization, evolutionary computation, robotics, artificial intelligence, intelligent control, feedback control of dynamic systems, game theory, and synthetic biology.