Building Safe and Stable DNN Controllers using Deep Reinforcement Learning and Deep Imitation Learning

Abstract

Cyber-physical systems (CPSs) with controllers built using deep neural nets and reinforcement learning (DRL) have become increasingly used in the functioning of our society. How to assure the correctness such as the safety and stability of these DNN controllers is extremely important and remains a major research challenge. This paper presents an approach to build safe and stable DNN controllers using DRL and deep imitation learning (DIL). An initial DNN controller is built using DRL, which is used to bootstrap a behavior preserving target DNN controller with safety and stability guarantees via DIL. We have applied this approach in successfully building safe and stable DNN controllers of a simplified airplane pitch control system.