Deep Reinforcement Learning Approach to QoE-Driven Resource Allocation for Spectrum Underlay in Cognitive Radio Networks

This paper presents a deep reinforcement learning-based technique for cognitive radio underlay dynamic spectrum access (DSA) that performs distributed joint multi-resource allocation to satisfy the primary link interference constraint and to maximize the secondary network performance, measured through the Mean Opinion Score (MOS) metric. The use of MOS as performance metric enables seamless integrated resource allocation of dissimilar traffic. The resource allocation problem is solved by utilizing a Deep Q- Network (DQN) algorithm, an advanced deep reinforcement learning approach, and a neural network to approximate the Q action-value function. Moreover, the learning process is improved by incorporating transfer learning to the learning procedure. Simulation results show that transfer learning reduces the number of iterations for convergence by approximately 25% and 72% compared to the DQN- algorithm without utilizing transfer learning and standard Q- learning, respectively.