Learning aided dynamic spectrum access in decentralized networks with unknown number of secondary users

Abstract

Opportunistic spectrum access (OSA) in a decentralized network is a challenging problem since each unlicensed user (i.e. cognitive radio (CR)) needs to characterize frequency bands as per the occupancy statistics of multiple licensed users. In dynamic and heterogeneous networks, all CRs may not be active simultaneously and can leave or join the network at any time. This makes OSA more challenging especially in the decentralized network where active CRs do not have any knowledge of other CRs in the network. In this thesis, a new decision making policy (DMP) using online learning algorithms has been proposed for characterization of frequency bands and orthogonalization of CRs into different but optimal bands. The proposed DMP, when implemented at all CRs in decentralized network, leads to an order-optimal policy. There are two main underlying strategies 1) learning the sub-band availability, which helps to reduce the long waiting time due to occupancy by primary user (PU), and 2) learning the efficiency of rank, which helps in reducing collision among the secondary users (SUs). For the fixed number of SUs, the loss in throughput decreases by 63% in case of 4 SUs and 46.6% in case of 6 SUs compared to existing state-of-art DMPs, Musical Chair (MC). For varying number of SUs, the loss in throughput decreases by 81.5% in case of 4 SUs and 84.8% in case of 6 SUs compared to MC.