A Sequential Experience-driven Contextual Bandit Policy for MIMO TWAF Online Relay Selection

In this work, we derive a sequential experience-driven contextual bandit (CB)-based policies for online relay selection in multiple-input multiple-output (MIMO) two-way amplify-and-forward (TWAF) relay networks, where the relays are provided with quantized imperfect channel gain information. The proposed CB-based policy acquires information about the optimal relay node by resolving the exploration-versus-exploitation dilemma. In particular, we propose a linear upper confidence bound (LinUCB)-based CB policy, and an adaptive active greedy (AAG)-based CB policy that utilizes active learning heuristics. With simulation results, we show that the proposed CB-based policies can reduce the feedback overhead by a factor of eight and time-cost by 70% while outperforming the best conventional Gram-Schmidt (GS) algorithm.