Contextual Bandit-Based Channel Selection for Wireless LANs with Interference-Driven Feature Extraction.

This paper proposes a radio channel selection algorithm based on a contextual multi-armed bandit (CMAB) for a wireless local area network (WLAN) environment, in which the access probability of each access point (AP) and the throughput model are not given in advance. The problem to be considered inherently involves the exploration to obtain the knowledge of the throughput distribution, in which a realized value is observed only after attempting to select each channel. This can be formulated as a multi-armed bandit (MAB) problem; particularly, we focus on the usefulness of the surrounding channel allocation information as the side information and determine that CMAB is appropriate. However, directly applying common CMAB algorithms to the such problems can lead to the lack of learning efficiency when the number of contexts is large. To reduce the computational complexity of the CMAB algorithms, feature extraction is designed by focusing on interference with neighboring and same-channel APs after channel selection of a target AP, which also contributes to the learning efficiency. To learn the optimal channel efficiently, this study investigates the most efficient method among the typical CMAB algorithms, including epoch-greedy, LinUCB, and Thompson sampling. The simulation results reveal that the algorithm based on JointLinUCB learns most efficiently under the environment where the access probabilities of the APs are extremely different.