Robust Control of Linear Stochastic Systems with Fully Observable State

Abstract

We consider a multidimensional linear system with additive inputs (control) and Brownian noise. There is a cost associated with each control. The aim is to minimize the cost. However, we work with the model in which the parameters of the system may change in time and in addition the exact form of these parameters is not known, only intervals within which they vary are given. In the situation where minimization of a functional over the class of admissible controls makes no sense since the value of such a functional is different for different systems within the class, we should deal not with a single problem but with a family of problems. The objective in such a setting is twofold. First, we intend to establish existence of a state feedback linear robust control which stabilizes any system within the class. Then among all robust controls we find the one which yields the lowest bound on the cost within the class of all systems under consideration. We give the answer in terms of a solution to a matrix Riccati equation and we present necessary and sufficient conditions for such a solution to exist. We also state a criterion when the obtained bound on the cost is sharp, that is, the control we construct is actually a solution to the minimax problem. 1. Introduction. The classical stochastic control theory deals with a stochastic system in which the uncertainty is of exogenous type and is described by a stochastic process with known characteristics. In addition to the exogenous stochastic process the dynamics of such a system depends on the control functional (policy) which can be chosen within an a priori known class. Usually there is a cost associated with each control functional. The objective is to find the minimal cost and the minimizing (optimal)