PhD Dissertation Proposal by JIANG Hao | Reward Penalty and Value Decomposition for Reinforcement Learning with Constraint

ABOUT THE TALK

Reinforcement learning with constraints has become essential for solving decision-making tasks subject to various constraints, especially in complex multi-agent environments. This talk introduces two CRL methods: Reward Penalty and Value Decomposition. The Reward Penalty approach ensures safety by applying penalties to unsafe policies, guiding actions to remain within specified constraints. Additionally, Value Decomposition enables agents to work together efficiently by decomposing joint actions, making it effective in large-scale multi-agent settings. Applications include multi-agent ride-pooling, where efficient resource allocation and flexible pickup/drop-off points are critical for optimal performance.

ABOUT THE SPEAKER

Jiang Hao is a PhD candidate in Computer Science at School of Computing and Information Systems, supervised by Prof. Pradeep VARAKANTHAM. His research interests lie in the areas of reinforcement learning with constraints, with a particular focus on developing solutions using reward penalty and improving value decomposition to address complex multi-agent and multi-constraint problems.