Dimmer: Self-Adaptive Network-Wide Flooding with Reinforcement Learning
share
In low-power wireless networks, Synchronous transmissions (ST) protocols provide high reliability and energy-efficiency in the presence of little or no interference, while dependable ST protocols provide high reliability in harsher environments, through the use of custom rules, fixed configurations and higher retransmissions. Yet, such dependable solutions often trade energy-efficiency for dependability, favoring wasting energy under normal conditions to survive highly-interfered episodes. We argue that, complementary to their dependability, ST protocols should be adaptive: their wireless stack should (1) tackle external environment dynamics and (2) adapt to its topology over time. We introduce Dimmer as a self-adaptive, all-to-all communication primitive. Dimmer builds on LWB and uses Reinforcement Learning to tune its flooding parameters and match the current properties of the medium. By learning how to behave from unlabeled traces, Dimmer adapts to different interference types and patterns, and is even able to tackle previously unseen interference. In addition, we share through Dimmer insights on how to efficiently design AI-based systems for constrained devices, and evaluate our protocol on two deployments of 18 and 48 resource-constrained sensor nodes (4 MHz CPU, 10 kB RAM), showing it improves reliability under WiFi interference and IEEE 802.15.4 jamming, while turning superfluous transmitters off in the absence of disturbances.
READ FULL TEXT
