Multi-Armed Bandits for Spectrum Allocation in Multi-Agent Channel Bonding WLANs

Abstract

While dynamic channel bonding (DCB) is proven to boost the capacity of wireless local area networks (WLANs) by adapting the bandwidth on a per-frame basis, its performance is tied to the primary and secondary channel selection. Unfortunately, in uncoordinated high-density deployments where multiple basic service sets (BSSs) may potentially overlap, hand-crafted spectrum management techniques perform poorly given the complex hidden/exposed nodes interactions. To cope with such challenging Wi-Fi environments, in this paper, we first identify machine learning (ML) approaches applicable to the problem at hand and justify why model-free RL suits it the most. We then design a complete RL framework and call into question whether the use of complex RL algorithms helps the quest for rapid learning in realistic scenarios. Through extensive simulations, we derive that stateless RL in the form of lightweight multi-armed-bandits (MABs) is an efficient solution for rapid adaptation avoiding the definition of broad and/or meaningless states. In contrast to most current trends, we envision lightweight MABs as an appropriate alternative to the cumbersome and slowly convergent methods such as Q-learning, and especially, deep reinforcement learning.