PhD Dissertation Proposal by BUI The Viet | Reliable Multi-Agent Policy Learning from Limited and Imperfect Feedback

ABOUT THE TALK

Learning to act as a team is challenging when guidance is limited or imperfect. Demonstrations may show useful behavior without explaining each agent’s contribution, while comparative judgments identify better outcomes without specifying every action. When learning relies on previously collected data, incomplete coverage further complicates the relationship between a learned policy and the behavior it will produce. Safety requirements add a separate need to account for the costs of collective decisions.

This dissertation investigates cooperative policy learning from demonstrations, trajectory-level preferences, and offline data with explicit reward and cost signals. It studies structured value representations that connect local decisions to team objectives, methods for extracting executable policies from learned values, and distribution-aware optimization under expected-cost constraints. These approaches provide complementary perspectives on how the available feedback can support effective coordinated behavior.

The research is evaluated in cooperative multi-agent benchmarks and simulated robotics tasks, with attention to task performance, data assumptions, and constraint satisfaction. By bringing the methods and their limitations into a common account, the dissertation aims to clarify how teams can learn from different forms of guidance and account for stated cost requirements. The resulting insights are intended to inform the design of reliable multi-agent systems that support human activities.

ABOUT THE SPEAKER

BUI The Viet is a PhD candidate in Computer Science at the School of Computing and Information Systems, Singapore Management University, advised by Assistant Professor MAI Anh Tien. Viet’s research focuses on multi-agent decision-making and reinforcement learning, particularly imitation learning, preference-based learning, offline reinforcement learning, and safety-constrained policy optimization. His work investigates how cooperative agents can learn effective and reliable policies from limited and imperfect feedback. His research has been accepted at conferences including NeurIPS, ICLR, and ICML.