Natural Policy Gradient Preserves Spatial Decay Properties for Control of Networked Dynamical Systems

We consider the distributed control of networked linear time-invariant systems. Previous work has established the spatial decay property of the centralized controller, which allows truncating the centralized controller to obtain a k-hop distributed controller with small performance loss. This paper makes a step further by showing a policy optimization approach, Natural Policy Gradient (NPG), preserves the spa-tial decay property of controllers. This enables “truncating” Natural Policy Gradient to directly learn a k-hop distributed controller.