A2.1 - Decentralized Reinforcement Learning for Adaptive Transmission Parameter Optimization of a LoRa Transceiver
- Event
- ETTC 2024 - European Test and Telemetry Conference
2024-06-11 - 2024-06-13
Nuremberg - Chapter
- ML & AI Session I
- Author(s)
- J. Gissing - Technische Universität Berlin, Berlin (Germany), C. Brockmann - Fraunhofer Insitute for Reliability and Microintegration (IZM), Berlin (Germany)
- Pages
- 22 - 31
- DOI
- 10.5162/ETTC2024/A2.1
- ISBN
- 978-3-910600-02-7
- Price
- free
Abstract
In wireless sensor networks (WSN), a large share of the energy demand arises from wireless communication, especially in wide area networks where transmission distances are at the scale of kilometers. Ensuring reliability of communication links while optimizing energy demand requires heterogeneous radio configurations throughout the network demanding for an automated process for identifying suitable transceiver settings in order to mitigate the effort of manual configuration during deployment. Furthermore, wireless links are susceptible to dynamic influences such as environmental conditions and interference from concurrent channel usage, rendering static radio configuration impractical. Therefore, autonomous organization and self-configuration of wireless communication networks, such as transmission parameter optimization, drastically reduce cost and effort for installation and maintenance of large-scale sensor systems. Such dynamic adaptive behavior can be achieved by local execution of decentralized methods that enable decision-making at the network edge, while also inherently offering advantages such as enhanced system robustness and scalability. In this work, we present a method that exemplifies this approach and experimentally evaluate its performance on real hardware. The adaptive algorithm optimizes the transmitter configuration of a LoRa transceiver by employing a model-free reinforcement learning approach based on an actor-critic setup using a parameterized stochastic policy and state-value function approximation. Experimental results show that the approach surpasses a standard approach in terms of long-term energy demand. Furthermore, the method’s capability of adapting to dynamic wireless channels is demonstrated.