A State-Space Approach to Stochastic Decentralized Control

We present some new approaches to synthesis of optimal decentralized control systems. The focus is on stochastic systems, and we address both linear dynamical systems and more general Markov decision processes. We give explicit formulae for optimal controllers for some specific decentralized information structures. These results expose the structure of such controllers, and in particular show the role of estimation and the required state-dimension of optimal controllers. We further discuss the relationship of our approach to well-known methodologies for the centralized case, including spectral factorization and dynamic programming.