A Distributed Deep Koopman Learning Algorithm for Control

Abstract: This paper proposes a distributed data-driven framework to address the challenge of dynamics learning from a large amount of training data for optimal control purposes, named distributed deep Koopman learning for control (DDKC). Suppose a system states-inputs trajectory and a multi-agent system (MAS), the key idea of DDKC is to assign each agent in MAS an offline partial trajectory, and each agent approximates the unknown dynamics linearly relying on the deep neural network (DNN) and Koopman operator theory by communicating information with other agents to reach a consensus of the approximated dynamics for all agents in MAS. Simulations on a surface vehicle first show that the proposed method achieves the consensus in terms of the learned dynamics and the learned dynamics from each agent can achieve reasonably small estimation errors over the testing data. Furthermore, simulations in combination with model predictive control (MPC) to drive the surface vehicle for goal-tracking and station-keeping tasks demonstrate the learned dynamics from DDKC are precise enough to be used for the optimal control design.