Bandwidth-Constrained Communication in Multi-Agent RL
A new paper on arXiv introduces SLIM, a minimal architecture for multi-agent reinforcement learning (MARL) that decouples communication from policy representation under bandwidth constraints. The authors propose a normalized per-agent bandwidth budget β to unify sparsity, rounds, and message dimension. This addresses performance degradation in applications like drone swarms for search-and-rescue, where reducing message size limits policy capacity. The work is from arXiv:2605.21085.
Key facts
- arXiv:2605.21085
- SLIM architecture decouples communication from policy latent space
- β is a normalized per-agent bandwidth budget
- Addresses bandwidth constraints in MARL
- Application: search-and-rescue with drone swarms
- Coupled bottleneck in shared latent representation causes performance loss
- Reducing message size limits policy capacity
- In-step co-optimization is mentioned
Entities
Institutions
- arXiv