Deep Reinforcement Learning based Channel Allocation for Channel Bonding Wi-Fi Networks

This paper presents Deep Reinforcement Learning (DRL)-based channel allocation algorithms for Wi-Fi networks with channel bonding capability. In particular, the proposed DRL algorithms allocate the primary channel and the maximal bonding bandwidth for each access point (AP). Existing DRL-based channel allocation algorithms assume a pre-known static interference model between APs, which cannot be accurately obtained in the hidden terminal scenario and the hidden channel scenario where APs have different sensing capabilities depending on the used channels. In contrast, our proposed DRL algorithm leverages the observed throughput as a reward to learn the interference relationship automatically, and implement centralized and distributed algorithms based on Proximal Policy Optimization (PPO) to learn and optimize channel allocation policies for improved performance. Simulation results show that the proposed methods outperform traditional methods in terms of network throughput in scenarios with hidden terminals and channels, and also perform well in scenarios with dynamic traffic loads. The proposed algorithms are more suitable for practical applications since no prior system knowledge is required.