Underlay Cognitive Radio Resource Management with Hybrid Meta-Loss Learning

Abstract

Cognitive Radio (CR) is an adaptable communication device driven by a Cognitive Engine (CE). A suitable machine-learning strategy can increase the learning potential of CE. This work proposes an underlay Resource Allocation (RA) framework for maximizing cognitive user Quality of Experience (QoE) while simultaneously protecting the primary network. A non-convex optimization problem is resolved using a combination of two algorithms: Meta Actor-Critic (MAC) loss and Deep Deterministic Policy Gradient (DDPG). The incorporation of the MAC neural network into RA improves the actor’s performance, which in turn boosts the entire system’s learning speed. The proposed framework is validated through a comparative analysis with four contemporary RA models: Q learning, hybrid, dueling DQN, and MAML. The simulation findings exemplify that the proposed method achieves convergence rapidly and improves reward value in a shorter amount of time than the other four extant RA models. Further, the impact of increasing system size on the distortion versus scalability trade-off is also highlighted.