Compact Representations for Adaptive Critic Designs

Missouri University Of Science And Technology, Rolla MO

Investigators

Abstract

ABSTRACT Compact Value Functions for Value-Based Control of Distributed Systems This project will extend the P1's work on a previous NSF grant on intelligent control for distributed systems. That previous grant started from recent results in value-based intelligent control, sometimes called approximate dynamic programming (ADP) or reinforcement learning (RL). Older forms of control typically guarantee stability, under certain theoretical assumptions, but do not focus on optimal performance. The new methods focus on optimal performance, without assuming linearity or perfect knowledge of system dynamics. By using new model-based versions of ADP or RI, and by developing value approximation schemes suitable for distributed systems, he was able to demonstrate new levels of performance in test problems in distributed tasks such as fluid flow control and missile interception. The new project is intended to be the first stage in addressing larger and more general distributed control problems, such as manufacturing processes, job scheduling, network traffic, etc. The first stage will be to explore more powerful value approximation schemes, in order to enhance the ability to control systems governed by partial differential equations, such as fluid flows encountered in aerospace, plasma and industrial applications, to the extent possible in a small-scale exploratory effort.