Distributed Real-Time and Fair Resource Allocation for 5G Dense Cellular Networks Based on Deep Reinforcement Learning

This article considers a 5G dense cellular network scenario where wireless channels are dynamically changing. To address the resource allocation problems in 5G dense networks, a distributed real-time and fair resource allocation algorithm based on a single deep Q-network (DQN) and multiple local deep neural networks (DNNs) architecture (DRFRA-SDML) is proposed. Specifically, each downlink is modeled as an agent. Each downlink is equipped with a local DNN, allowing each agent to input locally observed information into the local DNN to select spectral subband and transmission power in real time. In the core network, the global experience replay buffer is utilized to collect local experiences collected by all local agents to train a global weight vector of the train DNN, which is shared by all local DNNs. This ensures that the computational complexity of each local agent does not depend on the size of the cellular network, resulting in excellent scalability. To achieve fairness between user devices, a reward function based on fairness has been designed. The simulation results show that, while ensuring fairness between user devices, DRFRA-SDML significantly outperforms Ideal fractional programming for fair resource allocation, fractional programming for fair resource allocation with a slot delay, and Random in both sum rate and real-time performance.