Safe Reinforcement Learning (SRL) Using Control Barrier Functions
This talk presents a method to learn barrier-certified safe controllers for safety-critical systems while providing an optimal performance with the focus on reinforcement learning approach. The problem describes designing optimal controllers for systems with unknown dynamics through the interaction while safety specifications of the system such as state constraints must be satisfied. We first start by reviewing the basics of reinforcement learning in control such as the overall framework, Bellman equation, actor/critic approximations and sequential improvement of controller by means of reducing the prediction error. Then, different types of control barrier functions and their application for restricting the states of the system within a desired safe region and therefore safety guarantee are discussed. The safe reinforcement learning problem is then formulated by means of control barrier functions to have a safe performance. Safety, stability and optimality of the proposed method are discussed and finally the off-policy reinforcement learning algorithm to implement the proposed method is presented.