Deep Reinforcement Learning Based Dynamic Resource Allocation in 5G Ultra-Dense Networks

Abstract

The rapid development of Internet of things (IoT) technology has promoted the densification of network infrastructure. Ultra-dense networks (UDN) will become a key technology in 5G networks. We investigated the dynamic resource allocation problem over 5G UDN. We considered the energy efficiency (EE) and spectral efficiency (SE) of the network. Therefore, the resource allocation problem at different moments was expressed as a joint optimization problem. Considering the dynamic nature of the environment, the EE and SE were dynamically weighted. In order to guarantee the long-term performance of UDN system, the joint optimization problem was described as a markov decision process (MDP). In view of the fact that the densification of the network makes the space explosion of MDP and makes it difficult to solve by traditional methods, the dueling deep Q network (Dueling DQN) method was proposed. Simulation results showed that compared with traditional Q-learning and DQN, this algorithm has obvious performance improvement.