LLM-Driven Communication Protocol Boosts Multi-Agent RL
Researchers propose LMAC, a novel framework that uses an LLM's reasoning to design a communication protocol for cooperative multi-agent reinforcement learning. LMAC enables agents to reconstruct the underlying state accurately and uniformly, iteratively refining the protocol with an explicit state-awareness criterion. Experiments on diverse MARL benchmarks show improved state reconstruction and substantial performance gains over prior methods.
Key facts
- LMAC leverages LLM reasoning to design communication protocols
- Protocol enables all agents to reconstruct underlying state
- Iterative refinement using state-awareness criterion
- Experiments on diverse MARL benchmarks
- Outperforms prior communication baselines
Entities
—