Multiplay Multiarmed Bandit Algorithm Based Sensing of Noncontiguous Wideband Spectrum for AIoT Networks

To bring large-scale artificial intelligence of things (AIoT) to reality, wireless networks need intelligence to identify resources in a limited shared noncontiguous spectrum. In this article, we address this challenge via a sub-Nyquist sampling-based wideband spectrum analyzer deployed in the AIoT gateway. The noncontiguous nature demands learning the channel occupancy. However, the identification of channel status can fail when the number of busy channels in a selected subset is higher than the number of analog-to-digital converters,
$K$
. We model this subset selection problem as multiplay multiarmed bandit. First, we demonstrate the learnability of such a problem via a learning algorithm with a subset size of
$K$
(no sensing failure). For wideband sparse spectrum, we extend this algorithm using a novel subset size estimation approach to identify the optimal subset that gives the best possible throughput and could have a size potentially larger than
$K$
. These algorithms are mapped on the system-on-chip, and in-depth performance analysis demonstrates their superiority over state-of-the-art approaches.