An improved SPSA algorithm for stochastic optimization with bound constraints
Authors: | Popovic Dobrivoje, United Technologies Research Center, United States Teel Andrew R., University of California Santa Barbara, United States Jankovic Mrdjan, Ford Motor Company, United States |
---|
Topic: | 1.4 Stochastic Systems |
---|
Session: | Control, Estimation and Analysis of Stochastic Systems |
---|
Keywords: | Stochastic approximation, algorithms, real-time, parameter optimization, constrainedparameters, engine efficiency |
---|
Abstract
We show that the Simultaneous Perturbation Stochastic Approximation (SPSA) algorithmwith projection may exhibit slow convergence in constrained stochastic optimizationproblems when the optimum is situated on the constraints. The cause of the slowconvergence is a geometric interaction between the projection operatorand the SPSA gradient estimate. The effect of this interaction can be described as ``bouncing ofiterates against the constraints.'' We describe this on two lowdimensional noise-free examples, and present a new algorithm thatdoes not exhibit the bouncing effect and the consequent slowconvergence.