| |
 | | Reinforcement Learning Methods for Safety-Aware Decision-Making |  | BRAHMANAGE Janaka Chathuranga Thilakarathna PhD Candidate School of Computing and Information Systems Singapore Management University FULL PROFILE |
Research Area - Artificial Intelligence & Data Science
- Decision Making & Optimization
Dissertation Committee | Advisor: | | | Members: | | | | | | External Members: | Ze GONG, Assistant Professor, Chinese Academy of Sciences, Shenzhen Institute of Advanced Technology |
| | Date 25 November 2026 (Wednesday) Time 2:00pm – 3:00pm Venue Meeting room 4.4, Level 4 School of Computing and Information Systems 1, Singapore Management University, 80 Stamford Road, Singapore 178902 Please register by 24 November 2026. We look forward to seeing you at this research seminar.
|
| ABOUT THE TALK Reinforcement learning (RL) has achieved strong results across robotics, algorithm discovery, and classical games. However, deploying RL in safety-critical domains requires policies that satisfy safety constraints in addition to maximizing reward. This thesis addresses three settings for enforcing such requirements: action constraints, which restrict the set of admissible actions at every step; cumulative cost constraints, which bound the expected cost accumulated over a trajectory; and state-wise constraints, which require every state the agent visits to be safe. Projection-based approaches to action constraints can be computationally expensive and hinder policy learning when proposed actions fall far outside the feasible action space. To address action constraints, we present FlowPG, which learns a mapping to the feasible action space, reducing reliance on computationally expensive action corrections. However, a key challenge in FlowPG is the generation of feasible samples for complex constraints. We therefore present CV-Flow, which trains the flow model using a constraint violation signal. For cumulative cost constraints, we present Budget-Conditioned Reachability RL (BCRL), which defines a reachability set conditioned on the remaining safety budget. This formulation decouples reward maximization from the cumulative cost constraint, resulting in non-adversarial optimization. For state-wise constraints, we learn a safety value function from offline data to estimate the set of state-action pairs from which the agent can remain safe indefinitely. This accounts for actions that appear safe immediately but can lead to unavoidable future violations. We provide theoretical analysis and empirical evaluation on a diverse set of tasks, including continuous control, offline safe-RL, and maritime navigation. | ABOUT THE SPEAKER BRAHMANAGE Janaka is a fourth-year PhD candidate in Computer Science at the School of Computing and Information Systems, Singapore Management University (SMU), advised by Prof. Akshat Kumar. He holds a B.Sc. Eng (Hons.) in Computer Science & Engineering with First Class Honours from the University of Moratuwa, Sri Lanka. His research focuses on safe reinforcement learning and imitation learning, with publications at NeurIPS, AAAI, IJCAI, ICAPS and AAMAS. He received the SMU Presidential Doctoral Fellowship (2024–2026) and was named to the SCIS Dean’s List (AY2026–27). He recently completed an internship at Huawei Singapore Research Centre, working on agentic search and tool-use agents for Celia. Before his PhD, he worked as a Research Engineer at SMU on transport optimisation and data mining. He also co-founded two tech startups in Sri Lanka. Outside research, he enjoys running, cycling and playing the ukulele. |
|