A Sensitivity-Based View of Learning and Optimization
Xiren Cao
Department of Electrical and Electronic Engineering
Hong Kong University of Science and Technology
Wednesday, October 26, 2005
4:30 - 5:45 PM
Terman Engineering Center, Room 453
Abstract:
The subject of learning and optimization has been attracting wide
attentions from researchers in many disciplines including control
systems, operations research, and computer science. Areas such as
perturbation analysis (PA), Markov decision processes (MDPs),
reinforcement learning (RL), and adaptive control (AC), share a common
goal: to optimize the system performance. In this talk, we present a
sensitivity-based view of learning and optimization of stochastic
dynamic systems.
1. We show that the sensitivity-based view provides a unified
framework for learning and optimization, and many existing results in
the area can be derived from the two types of performance sensitivity
equations, one for performance derivatives and the other for
performance differences.
2. We study the Markov decision problem by using the performance
difference equations and show that the long-run average performance-,
the bias-, and the nth-bias- optimization problems can be easily and
intuitively solved by this approach with no discounting or Laurent
expansion. This simplifies the proof of Veinott¢s early results, and
we hope that it can help to popularize these important works.
3. We introduce the event-based learning and optimization approach, which may be applied to problems that cannot be solved by the standard MDPs.
Operations Research Colloquia: http://or.stanford.edu/oras_seminars.html