WorksMathematics
Action-Dependent Policies for Multi-Agent Linear Quadratic Regulation
Abstract
The linear quadratic regulation (LQR) problem over a multi-agent system usually requires a distributed feedback gain that is structurally constrained by the network topology, incurring an inevitable suboptimality gap. Existing studies either identify conditions on the system and topology that enable convex relaxation or seek a stationary point via policy gradient algorithms. In this paper, we reduce the suboptimality gap by adopting action-dependent policies (ADPs), where the control input of each agent depends on not only states of neighbor agents, but also control inputs of some other agents. We show that although exhibiting a distributed structure, ADP can enlarge the feasible set of feedback gains and even represent centralized optimal policies under suitable conditions. However, this comes at the price of losing coercivity of the cost function and possible failure of convergence for the projected gradient flow (PGF). To overcome this difficulty, we propose a partially regularized formulation, which guarantees convergence of PGFs, and forges a connection between standard multi-agent LQR and its action-dependent counterpart. Sufficient conditions for local convergence to globally optimal policies under PGFs regardless of regularization are also established. Finally, numerical examples are presented to validate the performance of ADP.