March 2009 Meeting:
Wednesday, March 11, 2009

Carnegie Mellon University, Silicon Valley (directions:   https://sv.cmu.edu/who_we_are/visitor)

FREE

Control on Landscapes with Local Minima and Flat Regions: A Simulated Annealing and Gain Scheduling Approach

Dr. Abe Ishihara   

Convergence of the backpropagation algorithm and its variants highly depend on the shape of the landscape. It is well known that multilayer neural networks employing sigmoid-like nonlinearities in the hidden layers result in landscapes with local minima. Another type of local minima, in which we term "flat regions" arise due to the fact that the derivative of sigmoid-like activation functions tend exponentially to zero. In this talk we propose a new control method to deal with these issues. This method employs simulated annealing and a gain-scheduled learning rate for the trajectory tracking control of a nonlinear system, such as the multi-link robot manipulator. Using stochastic calculus, we are able to prove that this method results in a closed loop system that is semi-globally uniformly bounded in expected value. Furthermore, we illustrate on a simple example how the proposed method escapes local minima and flat regions, whereas a conventional neural network controller gets "stuck". Extensions to the control of the multi-link robot manipulator are discussed and simulations are presented.

Dr. Abe Ishihara:

Dr. Ishihara received the B.S. degree in Electrical Engineering from Rensselaer Polytechnic Institute, and the M.S. and Ph.D. degrees in Aeronautics and Astronautics from Stanford University in 1998, 2002, and 2008, respectively. Prior to coming to Stanford, he worked on the adaptive control of drug delivery systems, reconfigurable flight control with saturating actuators and Helicopter Active Control Technology (HACT). At Stanford, he worked on the neural network control of robot manipulators, asymptotic methods for the stability and instability analysis of nonlinear systems and simulated annealing for nonlinear control. Additionally, he developed methods for measuring human motor learning rates in unknown environments using surface electromyography. He was the recipient of the NIH National Research Service Award (NRSA) training grant in which he investigated models of human motor adaptation during arm reaching with robot manipulator induced perturbations.  

Online video of event.