Coordinating Coexisting Learning Agents in Shared Spectrum via Parameter Space Complementarity

Independently managed Low Power Wide Area Networks (LPWANs) such as LoRa and SigFox increasingly share crowded unlicensed spectrum. Recent systems push adaptation into the LoRa Network Server (LNS) and use reinforcement learning to tune transmission parameters such as channel, spreading factor, and power. When multiple deployments operate learning agents at their LNS and overlap in time and frequency, independent agents exposed to similar conditions can drift toward the same resource choices, creating persistent collisions and non stationary dynamics that slow or destabilize learning. We propose a lightweight coordination layer that exchanges only model level information. Each LNS periodically shares policy parameters and an optional compact performance summary, and a Collaborative Coordination Service synthesizes individualized updates that encourage complementary operating regions rather than global consensus. We describe a hypernetwork style coordination mechanism and a composite objective that balances local utility preservation with cross network diversity. Experiments on a small physical setup and representative NS-3 simulations show that parameter space coordination can improve reliability and reduce transmissions per delivered packet relative to uncoordinated deep RL agents, while requiring no raw traffic sharing and only periodic backhaul side updates. We also discuss practical safety, privacy, and robustness considerations for deploying such coordination infrastructure in open environments.