PhD Dissertation Defense by LOW Siow Meng | From Sparse Feedback to Sequential Decision-Making: Learning Safety Constraints with Weak Supervision

From Sparse Feedback to Sequential Decision-Making: Learning Safety Constraints with Weak Supervision

ABOUT THE TALK

Real-world decision-making often involves safety requirements that are implicit or difficult to define explicitly. Traditional reinforcement learning (RL) methods typically assume access to detailed cost functions and predefined safety constraints, which limits their applicability in practice. This dissertation presents a unified framework for learning safety constraints directly from high-level, weak supervision—such as binary feedback on entire trajectories—rather than relying on handcrafted specifications.

The work begins with an efficient planning algorithm for continuous environments with known dynamics, forming the foundation for subsequent developments. The core contributions then focus on algorithms that infer both Markovian and non-Markovian safety constraints from sparse supervision, including settings where the environment dynamics are unknown. Notably, the proposed frameworks enable agents to learn safe behaviours even with limited, coarse-grained trajectory-level feedback. 

Together, these methods bridge the gap between data-driven constraint learning and safe reinforcement learning, paving the way for deploying intelligent agents in environments where safety requirements are dynamic, implicit, or hard to formalize.

SPEAKER BIOGRAPHY

Low Siow Meng is a PhD candidate in Computer Science at the SMU School of Computing and Information Systems, where he conducted research under the supervision of A/P Akshat Kumar. He holds a Bachelor’s degree from the National University of Singapore and a Master’s degree from Imperial College London. Prior to joining SMU, he gained software R&D experience at technology companies including IBM and NEC. 

His research interests lie at the intersection of reinforcement learning, data-driven constraint learning, and safety in sequential decision-making. His doctoral work focuses on inferring safety constraints from weak supervision and developing safe learning algorithms for complex environments. His research has been presented at leading venues such as AAAI, ICAPS, and an ICLR workshop. 

Beyond research, Siow Meng enjoys running, following strategic board games, exploring new productivity tools, and reading about the intersection of AI and society.