Timezone: »

Stochastic Search and Optimization
James Spall

Mon Dec 03 03:30 PM -- 05:30 PM (PST) @ Emerald Bay A, Harveys Convention Center Floor (CC)

Stochastic search and optimization (SS&O) methods are widely used in many areas of computational science. Online algorithms, such as stochastic gradient descent, are a prominent example of SS&O. The speaker will discuss some general issues related to how SS&O contributes to the analysis and control of modern systems as a way of: (i) coping with inherent system noise, (ii) providing algorithms that are relatively insensitive to modeling uncertainty, and (iii) providing algorithms that are able to find a global solution from among multiple local solutions. As a specific example of SS&O, the speaker will discuss the simultaneous perturbation stochastic approximation (SPSA) algorithm for difficult multivariate optimization problems arising in stochastic systems. The essential feature of SPSA, which accounts for its power and relative ease of use in difficult multivariate optimization problems, is the underlying gradient approximation that requires only two objective function measurements regardless of the dimension of the optimization problem. This talk will focus on the basic ideas and motivation behind SPSA without dwelling on the mathematical details. As time permits, the speaker will also include some discussion on contrasts with other algorithms (genetic algorithms, simulated annealing, etc.) and will briefly discuss some recent advances in areas such as discrete optimization and adaptive (second-order) search with or without stochastic gradients.

Author Information

James Spall (The Johns Hopkins University)

James C. Spall is a member of the Principal Professional Staff at the Johns Hopkins Applied Physics Laboratory, a Research Professor in the JHU Department of Applied Mathematics and Statistics, and the Chairman of the Applied and Computational Mathematics Program within the JHU Engineering Programs for Professionals. Dr. Spall has published extensively in the areas of control systems and statistics and holds two U.S. patents for inventions in control systems, both licensed to U.S. companies. He is the editor and coauthor of the book Bayesian Analysis of Time Series and Dynamic Models (CRC Press) and the author of Introduction to Stochastic Search and Optimization (Wiley). Dr. Spall is one of the inaugural Senior Editors for the IEEE Transactions on Automatic Control and is a Contributing Editor for the Current Index to Statistics. He was the Program Chair for the 2007 IEEE Conference on Decision and Control and is a Fellow of IEEE.