MABAMS: Multi-Armed Bandit-Aided Mode Selection in Cooperative Buffer-Aided Relay Networks

In several practical networks the environment is changing fast; however, acquiring channel state information (CSI) continuously is often infeasible. In such cases, it is often feasible to infer statistical CSI at the transmitters (CSIT). In this paper, we study the mode selection problem for a cooperative network, consisting of a source, a butter-aided (BA) full-duplex (FD) relay, and a destination. In this setting, at every time frame, the network can operate in either FD mode (in different power levels), or, switch to half-duplex (HD) mode when the FD mode is not feasible. Aiming to choose the best mode of operation, a mode selection mechanism is proposed, named MABAMS, which makes use of a multi-armed bandit learning approach that exploits the acknowledgements/negative-acknowledgements (ACKs/NACKs) observations in order to extract useful information about the statistics of the channels. As a consequence, MABAMS avoids the need for CSI acquisition and exchange. We provide a performance evaluation in terms of average throughput, in order to demonstrate an interesting performance-complexity trade-off when compared to the case in which the channel statistics are known. In addition, we demonstrate significant performance improvements over the cases without any power adaptation.