Learning-Based Multi-Channel Spectrum Access in Full-duplex Cognitive Radio Networks with Unknown Primary User Activities

Abstract

Cognitive radio had been proposed as a methodology for overcoming the inefficiency of the conventional static allocation of the available spectrum in wireless communication networks. The majority of opportunistic spectrum access schemes in cognitive radio networks (CRNs) rely on the Listen-Before-Talk (LBT) model due to the half-duplex nature of conventional wireless radios. However, LBT su ers from the problem of high collision rates and low secondary user throughput if time is misaligned among the secondary users (SUs) and the primary users (PUs). This problem can be mitigated by leveraging full-duplex (FD) communications that facilitate concurrent sensing and transmission. This thesis considers the problem of optimal opportunistic multi-channel spectrum sensing and access using FD radios in the presence of uncertain primary user (PU) activity statistics. A joint learningand spectrum access scheme is proposed. To optimize its throughput, the SU sensing period has to be carefully tuned. However, in absence of exact knowledge of the PU activity statistics, the PU's performance may be adversely a ected. To address this problem, a robust optimization problem is formulated. Analysis shows that under some non-restrictive simplifying assumptions, the robust optimization problem is convex. The impact of sensing periods on the PU collision probability and the SU throughput are analyzed, and the optimal sensing period is found via convex optimization. An "\epsilon-greedy algorithm is proposed for use by the SU to learn the PUs' activity statistics in multichannel networks. It is shown that sublinear regrets can be attained by the proposed estimation and robust optimization strategy. Simulation studies demonstrate that the resulting robust solution achieves a good trade-o between optimizing the SU's throughput and protecting the PU.