세미나

📑Abstract:
Multi-agent reinforcement learning has gained attention as a method to build autonomous agents and has been applied to various systems. In cooperative multi-agent reinforcement learning (MARL), agents aim to achieve a common goal, such as defeating enemies in a competition or scoring a goal in a soccer game. While existing MARL algorithms are effective in relatively easy tasks, they still require significant learning time and often get trapped in local optima when faced with complex tasks. This leads to failures in learning paths toward such semantic goals or goal-reaching trajectories. To address these challenges, we introduce (1) semantic embedding and (2) memory-based incentive. Semantic embedding is used to construct memories, grouping similar experiences close together in the embedding space, enabling broader and more effective memory recall during training. Memory-based incentive, on the other hand, is about how these memories are used. It promotes transitions toward goal-reaching paths by correcting errors in value estimation. The combination of these two components enhances sample efficiency during training and facilitates better policy learning. In this talk, I will describe the key insights behind these methodologies and discuss potential future work and applications.👤Short Biography:
Hyungho Na is a postdoctoral researcher at the KAIST Institute for Robotics. His research interests include (multi-agent) reinforcement learning, mission planning and execution, representation learning, and their applications across various domains. From 2013 to 2021, he worked as a researcher and senior researcher at the Agency for Defense Development (ADD), where he contributed to the R&D of missile defense systems. He holds a B.S., M.S., and Ph.D. in Aerospace Engineering (with a concentration in multi-agent reinforcement learning), all from KAIST.