A Dueling Deep Recurrent Q -Network Framework for Dynamic Multichannel Access in Heterogeneous Wireless Networks

Abstract

This paper investigates a deep reinforcement learning algorithm based on dueling deep recurrent $Q$ -network (Dueling DRQN) for dynamic multichannel access in heterogeneous wireless networks. Specifically, we consider the scenario that multiple heterogeneous users with different MAC protocols share multiple independent channels. The goal of the intelligent node is to learn a channel access strategy that achieves high throughput by making full use of the underutilized channels. Two key challenges for the intelligent node are (i) there is no prior knowledge of spectrum environment or the other nodes’ behaviors; (ii) the spectrum environment is partially observable, and the spectrum states have complex temporal dynamics. In order to overcome the aforementioned challenges, we first embed the long short-term memory layer (LSTM) into the deep $Q$ -network (DQN) to aggregate historical observations and capture the underlying temporal feature in the heterogeneous networks. And second, we employ the dueling architecture to overcome the observability problem of dynamic environment in neural networks. Simulation results show that our approach can learn the optimal access policy in various heterogeneous networks and outperforms the state-of-the-art policies.