Optimal Tracking Of Nonlinear Discrete-time Systems Using Zero-Sum Game Formulation And Hybrid Learning

This paper presents a novel hybrid learning-based optimal tracking method to address zero-sum game problems for partially uncertain nonlinear discrete-time systems. An augmented system and its associated discounted cost function are defined to address optimal tracking. Three multi-layer neural networks (NNs) are utilized to approximate the optimal control and the worst-case disturbance inputs, and the value function. The critic weights are tuned using the hybrid technique, whose weights are updated once at the sampling instants and in an iterative manner over finite times within the sampling instants. The proposed hybrid technique helps accelerate the convergence of the approximated value functional to its actual value, which makes the optimal policy attain quicker. A two-layer NN-based actor generates the optimal control input, and its weights are adjusted based on control input errors. Moreover, the concurrent learning method is utilized to ease the requirement of persistent excitation. Further, the Lyapunov method investigates the stability of the closed-loop system. Finally, the proposed method is evaluated on a two-link robot arm and demonstrates promising results.