Data-based Optimal Control for Discrete-time Systems via Deep Deterministic Policy Gradient Adaptive Dynamic Programming

The model-free optimal control problem for discrete-time systems is considered in this paper by using deep deterministic policy gradient adaptive dynamic programming (DDPGADP) algorithm. The system data is obtained by using the off-policy learning and the control law is updated by policy gradient. The convergence of DDPGADP algorithm is verified by showing that the Q-function sequence is monotonically non-increasing and converges to the optimum. In order to implement this method, an actor-critic neural network structure is established by adopting the target network technology from deep Q-learning during the neural network training process. Finally, simulation examples are presented to verify the effectiveness of the proposed method.