Research Summary
My primary research objective addresses the need for trustworthy and safe Reinforcement Learning (RL) policies, particularly in safety-critical applications like autonomous vehicles, robots, and drones. Deep Reinforcement Learning (DRL) algorithms are used in these applications due to their impressive learning capabilities, but unfortunately, they come at the cost of limited interpretability. This institutes a trust deficit for their real-world deployment, as they can potentially lead to accidents that hinder societal acceptance. My research focuses on two critical aspects of achieving trustworthy and Safe RL (SRL). Firstly, we develop techniques to ensure RL agents learn without violating safety constraints, thus incurring a surge in training costs or fatal accidents. This involves constant runtime monitoring and devising new learning algorithms to ensure safety adherence. Secondly, we have explored methods to reduce the size and complexity of learned DRL policies while preserving their performance. Such reduced-sized policies are more interpretable and facilitate verification, enabling broader trust and acceptance of these technologies.
We have also developed a framework with Intel for generating targeted stimuli to boost simulation-based test coverage using machine learning models. At the same time, we have devised a strategy to find the root cause of an event from the traces of pre-silicon (RTL) hardware simulations. My research’s ultimate goal is to contribute to the widespread adoption of dependable and trustworthy ML-based systems that benefit society.