Multi-Agent Communication Optimized via Information Bottleneck and Vector Quantization
A novel framework designed for multi-agent reinforcement learning (MARL) systems tackles significant communication limitations found in practical robotics. By combining vector quantization with information bottleneck theory, this method facilitates selective and efficient communication that conserves bandwidth. It employs information-theoretic optimization to compress and discretize messages while maintaining essential task-related information. A gated communication system determines the necessity of communication based on the context of the environment and the states of the agents. Coordination task experiments reveal a 181.8% enhancement in performance compared to scenarios without communication and a 71.4% decrease in bandwidth consumption. Analysis of the Pareto frontier demonstrates superiority throughout the entire success-bandwidth trade-off. The paper can be accessed on arXiv (arXiv:2602.02035) and was introduced as a replace-cross update.
Key facts
- Framework combines information bottleneck theory with vector quantization.
- Achieves 181.8% performance improvement over no-communication baselines.
- Reduces bandwidth usage by 71.4%.
- Introduces a gated communication mechanism.
- Pareto frontier analysis shows dominance across success-bandwidth trade-offs.
- Targets multi-agent reinforcement learning systems in real-world robotics.
- Paper available on arXiv with ID 2602.02035.
- Announce type: replace-cross.
Entities
Institutions
- arXiv