Time- and Frequency-Domain Dynamic Spectrum Access: Learning Cyclic Medium Access Patterns in Partially Observable Environments

Abstract

Upcoming communication systems increasingly often tackle the spectrum scarcity problem through the coexistence with legacy systems in the same frequency band. Cognitive Radio presents popular methods for Dynamic Spectrum Access (DSA) that enable coexistence. Historically, DSA meant a separation solely in the frequency domain, while in recent years it has been extended through the dimension of time, by employing Machine Learning to learn semi-deterministic and cyclic medium access patterns of the legacy system that are observed through channel sensing. When this pattern is learnable, then a new system can utilize a neural network and predict future medium accesses, thus steering its own medium access. We investigate this novel and more fine-grained version of DSA, propose a predictor and show its capability of reliably predicting future medium accesses of a legacy system in an aeronautical coexistence scenario. We extend the predictor to the case of partial observability, where only a narrowband receiver is available, s.t. observations are limited to a single sensed channel per time slot. In particular, we propose a custom loss function that is tailored to partially observable environments. In the spirit of Open Science, all implementation files are released under an open license.

Sebastian Lindner

Doktorand

Maschinelles Lernen für Kommunikationsnetze, Verteilte Algorithmen, Kanalzugriff bei ad-hoc Netzen, Reinforcement Learning, Zuverlässigkeitsanalyse, Modellierung

Daniel Stolpmann

Doktorand & MLE-Webmaster

Machine Learning for Communication Networks, Low Latency & High Throughput Communication, Mobile Edge Computing, Satellite Communication, Network Coding

Andreas Timm-Giel

Professor & Präsident

Communication Networks, Performance Evaluation, Modelling, Machine Learning for Communication Networks