QoE Driven Resource Allocation in Massive IoT: A Deep Reinforcement Learning Approach

Dense deployment of various Internet of Things (IoT) equipment in cellular will lead to a sharp shortage of frequency resources. Consequently, how to make efficient utilization of radio resources in massive IoT scene is a brand-new challenge. In this paper, we present a novel deep reinforcement learning approach for massive IoT networks that performs centralized joint resource allocation to satisfy the link interference constraint and to maximize the quality of user experience (QoE) which is measured through utility based Mean Opinion Score (MOS). Different from the traditional resource allocation algorithm, we model the problem as a joint optimization resource allocation problem and use a neural network embedded reinforcement learning algorithm to obtain the optimal solution. In addition, we utilize a new reinforcement framework called actor-critic (AC) to overcome the problem of continuous power allocation that cannot be accomplished in traditional reinforcement learning algorithms. The simulation results show that our proposed algorithm has good convergence under different conditions. Furthermore, QoE is consistent with the optimal value under our algorithm, and the average QoE can reach a higher level.