Data-Driven Adaptive Dynamic Programming for Nonlinear Systems with State and Input Constraints

In this paper a novel data-driven adaptive dynamic programming approach for nonlinear systems with state and input constraints is developed. The basic idea is to transform state constraints into input constraints by introducing a barrier index function. Then a Q-learning approach is designed to solve the optimal control problem considering only input constraints. After that, a data-driven constrained value iteration approach is developed to obtain the optimal control policy. It is shown that the iterative action-value function converges to the optimal action-value function. Finally, a torsional pendulum system is used to illustrate the proposed approach.