Stochastic Optimization and Control
625.743
Course Description and Philosophy
Analysts, engineers, and managers are often faced with making tradeoffs
between different factors in order to achieve desirable outcomes. Choosing
these tradeoffs in the "best" way is the essence of optimization and learning.
Stochastic search and optimization plays an increasing role in 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.
This course introduces the fundamental issues in stochastic search and optimization with special emphasis on cases where classical deterministic techniques (linear and nonlinear programming, etc.) are inappropriate. These cases include many important practical problems, which will be discussed throughout the course (e.g., neural network training, global optimization, simulation-based optimization, nonlinear control, image processing, discrete-event systems, experimental design, etc.). Discrete and continuous optimization problems will be considered. Algorithms for global and local optimization problems will be discussed. Approaches such as random search, recursive least squares, Kalman filtering, stochastic approximation, simulated annealing, evolutionary and genetic algorithms, machine learning, and multiple comparisons are discussed.
The course will cover the relationships among algorithms and provide some guidance about which algorithms are appropriate for which problems. The emphasis will be on rigorous analysis and interpretation, with the aim of giving the students the tools to help sort through competing and conflicting claims in the literature. A student who successfully completes this course will be in a position to implement the algorithms and/or critically evaluate the implementations of others in a wide variety of practical problems.
The grade for the course will be based on a combination of performance on homework and final project. Homework will be assigned most weeks and will be due the following class. There are no in-class tests.
This course is co-listed in the Applied and Computational Mathematics and Electrical Engineering Programs of JHU Part-Time Engineering; it may serve as an elective in other JHU programs subject to adviser approval.
Subjects to be covered
1. Basic issues in search and optimization; stochastic vs. deterministic
methods; no free lunch theorems, summary of classical methods of optimization
and their limitations; review of multivariate analysis and matrix algebra
(Chapter 1 and Appendix A).
2. Introduction to random search; Monte Carlo methods; discussion
of probabilistic convergence and review of basic statistical tests (Chapter
2 and Appendices B and C).
3. Nonlinear simplex method; continued review of probabilistic
convergence and basic statistical tests (Chapter 2 and Appendices B and
C).
4. Recursive methods for linear systems: least mean squares,
recursive least squares, and Kalman filter (Chapter 3).
5. Root-finding and gradient-based stochastic approximation
(Robbins-Monro) (Chapter 4).
6. Stochastic gradient-based stochastic approximation (Chapter
5).
7. Gradient-free stochastic approximation with emphasis on the
finite-difference (FDSA) method (Chapter 6)
8. Simultaneous perturbation stochastic approximation (SPSA):
efficiency and theoretical analysis; implementation (link
to supplementary information on SPSA)
9. Simulated annealing and related methods (Chapter 8 and Appendix
E).
10. Evolutionary computation with emphasis on genetic algorithms
(Chapter 9).
11. Evolutionary computation (continued): evolution strategies
and theory for genetic algorithms (Chapter 10)
12. Machine learning via temporal difference methods (Chapter
11)
13. Statistical multiple comparisons for selecting the best
of several options (discrete optimization) (Chapter 12)
14. Student presentations of final reports and discussions of
other projects
Prerequisites
Students should come into the class with a working knowledge of probability
and statistics at the beginning graduate level (e.g., 625.403) and
knowledge of multivariate calculus and basic matrix algebra. Previous optimization
experience is not required. Some assignments will require the use of a
computer; the student is free to use any programming language with which
he/she is comfortable (the instructor will use MATLAB and will occasionally
pass out MATLAB code). Students are encouraged to contact the instructor
for additional information.
Instructor
James C. Spall joined the Johns Hopkins University Applied Physics
Laboratory in 1983 and was appointed to the Principal Professional Staff
in 1991. He is also Chair of the JHU Applied and Computational Mathematics
Program. Dr. Spall has published over 100 articles in the areas of statistics
and control and holds two U.S. patents. For the year 1990, he received
the Hart Prize as principal investigator of the most outstanding Independent
Research and Development project at JHU/APL. Among other appointments,
he is an Associate Editor at Large for the IEEE Transactions on Automatic
Control and a contributing editor for the Current Index to Statistics.
He is the editor and co-author for the book Bayesian Analysis of Time
Series and Dynamic Models (Marcel Dekker) and is the author of the
course textbook Introduction to
Stochastic Search and Optimization (Wiley). Dr. Spall has received
numerous research and publications awards and is a Fellow of IEEE, a member
of the American Statistical Association, and a Fellow of the engineering
honor society Tau Beta Pi. The perspective Dr. Spall will bring to the
course balances the theoretical aspects of the field with the practical
requirements for using search and optimization algorithms for solving real
problems.
E-mail the instructor.
Course Section, Location, and Time
| 625.743.31 | Applied Physics Laboratory (K4) | Thursday | 4:30-7:10 |
Computer Lab Requirements
Some homework assignments require the use of a computer. Students
are free to use whatever system and software is convenient. The instructor
uses PC-based MATLAB.
Textbook
The textbook for this course is Introduction
to Stochastic Search and Optimization by J. C. Spall (Wiley,
2003) (ISBN: 0-471-33052-3). The text will be available approximately
midway through the course In the interim, students will receive handouts
of relevant material from the book. A comprehensive collection of
supplementary reference literature is on reserve in the Gibson Library
at APL. These papers are largely optional reading, but are provided to
indicate other perspectives and to provide more detailed information on
certain subjects than are available in the textbook. Many of these
papers are at a more advanced level than the material in the course. Some
students may find these papers useful in carrying out the final class project.
Return to Part-Time Engineering
Return to Whiting School of Engineering
Return to JHUniverse
Applied
Math