HiComm: Hierarchical Communication for Multi-agent Reinforcement Learning

Recent research introduces three advancements in reinforcement learning. HyPOLE utilizes HyperLTL and hyperproperties to guide Multi-Agent Reinforcement Learning under partial observability. HiComm implements a plug-in communication module that grounds messages in hierarchical observations to improve cooperative MARL. Additionally, a new offline RL framework for fluid controls employs a sensor position-conditioned architecture to enable data-driven policy extraction, reducing the computational costs and real-time interaction requirements typically associated with classical online RL approaches in engineering.