Learning Through Imitation by Using Formal Verification

Avraham Raviv, Eliya Bronshtein, Or Reginiano, Michelle Aluf-Medina, Hillel Kugler

Abstract: Reinforcement-Learning-based solutions have achieved many successes in numerous complex tasks. However, their training process may be unstable, and achieving convergence can be difficult, expensive, and in some instances impossible. We propose herein an approach that enables the integration of strong formal verification methods in order to improve the learning process as well as prove convergence. During the learning process, formal methods serve as experts to identify weaknesses in the learned model, improve it, and even lead it to converge. By evaluating our approach on several common problems, which have already been studied and solved by classical methods, we demonstrate the strength and potential of our core idea of incorporating formal methods into the training process of Reinforcement Learning methods.