Joint Channel and Power Allocation in WLAN based on Sequential Deep Reinforcement Learning

In dense wireless local area networks (WLANs), users may experience unsatisfactory services owing to severe contention and interference caused by channel sharing among neighboring access points (APs). We propose a deep reinforcement learning (DRL)-based resource allocation (RA) scheme that can practically tackle the performance degradation problem by allocating appropriate channel and transmit (tx) power to each AP. To compensate for insufficient computing power of APs to train deep neural networks (DNN), the proposed scheme takes a hybrid approach of centralized training and distributed execution. Since the channel environment of neighboring APs can be very similar, when all APs determine their own channels at once, it is feasible for multiple neighboring APs to select the same channel and this can incur the performance degradation of the entire network. To avoid this phenomenon, the proposed scheme makes a DRL-based sequential decision, i.e., one AP at a time in a round-robin order determines its channel and tx power. Furthermore, the proposed scheme uses local observations that can be gathered from legacy 802.11 APs and stations (STAs), which may be a useful property because the local observations can be obtained even when there coexist uncontrolled APs in the environment and no additional functions need to be implemented in the STAs. We evaluate the performance of the proposed scheme using the average delay and the power consumption as metrics under unsaturated traffic condition. The experimental results show that the proposed scheme achieves better delay performance while consuming much lower power than comparison schemes in dense WLAN scenarios.