Mobility-Aware Centralized Reinforcement Learning for Dynamic Resource Allocation in HetNets

Heterogeneous networks (HetNets) can improve resource efficiency and coverage range in cellular networks to meet the growing demand for wireless data rate. The main challenges faced by HetNets are load balancing and interference coordination, which needs to be addressed by effective user association and resource allocation (UARA) methods. In this paper, we propose a mobility-aware centralized reinforcement learning (MCRL) framework in order to achieve global optimality of dynamic resource allocation. A centralized agent is defined to select the values of the hyper parameters for UARA according to the real-time status of all users in HetNets. Besides, the state of the art Actor-Critic technique is employed in the training process to guarantee the convergence and performance of the agent's policy. Simulation results demonstrate the effectiveness of the proposed method and show the performance gain under different user distributions.