Sufficient Statistics for Optimal Decentralized Control in System of Systems

Abstract

Research in multi-agent systems has mostly focused on heuristic/semi-heuristic methods for control, which lack in robustness and generalizability. Control theoretic techniques guarantee stability (and often optimality), but the results are limited in scope. Hence, there is a need to design intelligent control techniques as a function of sub-system dynamics, network structure and control/decision processes. We are developing S4C - a control theoretic framework for analysis and design of interacting robotic systems. We use “sufficient statistics” to generalize the separation principle - enabling decoupled optimal control and estimation. These techniques are applied to a missile guidance problem, demonstrating robustness to sensor/process noise. An agent-based simulation architecture has also been developed and used for studies. In addition, we use a verification and validation approach based on Gaussian process regression to test for cases where modeling assumptions are relaxed.