Safe Controller for Uncertain Nonlinear Systems using Model-based Reinforcement Learning

In this paper a reinforcement learning algorithm is presented to solve an infinite horizon optimal control problem for uncertain control-affine nonlinear systems under state constraint. We first develop a safe controller using time derivative of barrier Lyapunov function and then we use this to modify an actor-critic-identifier approach to safely learn optimal control policy. Furthermore, the safety analysis is studied. In the end, simulation example is given to prove the efficacy of the proposed method.