Contention-Driven Feature Extraction for Low-Regret Contextual Bandit-Based Channel Selection Dedicated to Wireless LANs.

To achieve low-regret learning in a radio channel selection for wireless local area networks (WLANs), we propose a contention-driven feature extraction (FE) scheme for a contextual multi-armed bandit (CMAB) algorithm. This study aims to learn the optimal WLAN channel online particularly in a scalable manner with respect to the number of APs and channels, which is accomplished by leveraging the context, i.e., channel allocation information. The proposed FE is designed by focusing on contention with neighboring and same-channel APs where the key idea is to consolidate contexts ignoring APs that are not connected to the target AP on the contention graph. The simulation results confirm that contention-driven FE enables a target AP to learn the optimal channel in a scalable manner for the number of APs and available channels and to have low regret using the CMAB algorithm.