Certified Reinforcement Learning

31^st March 2022, 13:00
Chao Huang

Abstract

There has been increasing interest in applying machine learning techniques, especially reinforcement learning, to control and general decision making. Due to the complexity of both machine learning algorithms and highly dynamic environments with significant uncertainties and disturbances, it is critical yet challenging to formally ensure the important properties of such learning-enabled systems, which hinders the adoption of machine learning in safety-critical scenarios, e.g., avionics systems and self-driving vehicles. In this seminar, I will introduce our recent work on certified reinforcement learning, i.e., providing safety and stability guarantees for reinforcement learning. First, we proposed safety verification techniques for a trained neural network based on Taylor model abstraction, i.e., design-then-verification. Then we integrated the verification techniques with the learning process, such that the properties can be automatically satisfied by learning itself, i.e., design-while-verify. We show by experiments that our approaches can significantly outperform the state of the arts regarding safety and stability. We believe our work is an important step towards a safe AI future.