An Algebraic Geometry Based Approach To Decentralized Control

Decentralized control has been one of the important problems in systems and control engineering. Computing an optimal decentralized controller for general linear systems, however, is known to be a very challenging task. In particular, designing an optimal decentralized controller in the standard framework of a linear system with quadratic cost and Gaussian noise is well known to be extremely hard even in very simple and small sized problems. Because of this fact, previous work has focused on characterizing several different classes of problems for which an optimal decentralized controller may be efficiently computed. The set of quadratically invariant problems is one of the largest known class of such problems. This dissertation provides a novel, general, and powerful framework for addressing decentralized control by introducing the idea of using rational elimination theory of algebraic geometry. We show that, in certain cases, this approach reduces the set of closed-loop maps of decentralized control to the solution set of a collection of linear equations. We show how to use these linear equations to find an optimal decentralized controller. We also prove that if a system is quadratically invariant then under an appropriate technical condition the resulting elimination set is affine. We further illustrate that our approach can be well applied to a strictly larger class of decentralized control problem than the quadratically invariant one by presenting a simple example: the example shows that there are problems which are not quadratically invariant but for which the resulting elimination description is affine.