Date of Award
Summer 9-5-2023
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
It is well-known that model correctness does not ensure the safe operation of systems that perform safety-critical functions. However, to ensure that potential hazardous events cannot occur, the model must be verified against its safety requirements, many derived during a systematic safety analysis. Typically, this process utilizes a complementary approach of formal verification and testing during the development process of complex systems. As these complex systems expand with the use of Reinforcement Learning (RL), safety considerations must be taken in the development of this technology. Although RL is becoming widely used across different industries, RL does pose some safety challenges. Few guidelines are available for developing a model for systems that demand a high level of safety confidence in an RL environment. The goal that I put forth is to increase safety confidence in systems that utilize RL by implementing an approach that incorporates industry-vetted safety guidelines and novel practices for developing safe RL models. The objective is to create a method that allows for the assessment of the safety risks associated with RL-based systems and determine their acceptability, providing a metric to measure this acceptability. The aim is to develop specific objectives and Level of Rigor (LOR) activities, which create guidelines for developers or project teams to increase the safety confidence in the model. These guidelines offer recommendations for enhancing the development of RL subsystems, facilitating the identification, evaluation, and mitigation of safety risks. It is essential to emphasize that these guidelines are designed to complement, rather than replace, the LOR activities established by traditional system safety standards to enhance the overall system’s safety confidence. The presented guidelines are demonstrated in a safety case to address both known unsafe and unknown unsafe hazards. For known unsafe scenarios, a deterministic analysis is utilized, while a rigorous development process is implemented to ensure safety in unknown unsafe scenarios, ensuring assurance and integrity in safety-critical functions. This approach involves a hazard analysis process to identify measures for risk mitigation (functional coverage) and LOR activities to ensure the high-quality model development (development coverage). The goal is to create guidelines that are comprehensive, contextually relevant, and easily understandable in addition to establishing a robust, accurate, reliable, and generalizable model.
Language
English (en)
Chair
Ramesh Agarwal