Learning Distributed Predictive Control via Spatial-Temporal Games

This article investigates learning predictive control framework for multiagent systems with unknown dynamics. Predictive control, which generates temporal control inputs, provides potential applications in observation loss scenarios. First, control causality is inversely extracted from time series analysis, and predictive control is characterized as sequential feedback control. Next, we focus on distributed communication and time consistency under the causality. Distributed predictive control is equivalently partitioned into spatial and temporal subgames, respectively. Spatial subgames achieve equivalence between global and local objectives through Nash equilibrium, while temporal ones force local control causality to achieve stability and optimality with time consistency. Furthermore, a multistep reinforcement learning algorithm is proposed for data-driven implementation, and dynamics knowledge is avoided with interactive data. The learning properties are discussed with theoretical proof and parameters selection. Finally, we pose numerical results to demonstrate effectiveness, and robotics experiments are also carried out to show potential advantage under observation loss scenarios.