Dimmer: Self-Adaptive Network-Wide Flooding with Reinforcement Learning

The last decade saw an emergence of Synchronous Transmissions (ST) as an effective communication paradigm in low-power wireless networks. Numerous ST protocols provide high reliability and energy efficiency in normal wireless conditions, for a large variety of traffic requirements. Recently, with the EWSN dependability competitions, the community pushed ST to harsher and highly-interfered environments, improving upon classical ST protocols through the use of custom rules, hand-tailored parameters, and additional retransmissions. The results are sophisticated protocols, that require prior expert knowledge and extensive testing, often tuned for a specific deployment and envisioned scenario. In this paper, we explore how ST protocols can benefit from self-adaptivity; a self-adaptive ST protocol selects itself its best parameters to (1) tackle external environment dynamics and (2) adapt to its topology over time. We introduce Dimmer as a self-adaptive ST protocol. Dimmer builds on LWB and uses Reinforcement Learning to tune its parameters and match the current properties of the wireless medium. By learning how to behave from an unlabeled dataset, Dimmer adapts to different interference types and patterns, and is able to tackle previously unseen interference. With Dimmer, we explore how to efficiently design AI-based systems for constrained devices, and outline the benefits and downfalls of AI-based low-power networking. We evaluate our protocol on two deployments of resource-constrained nodes achieving 95.8% reliability against strong, unknown WiFi interference. Our results outperform baselines such as non-adaptive ST protocols (27%) and PID controllers, and show a performance close to hand-crafted and more sophisticated solutions, such as Crystal (99%).