Spectrum prediction: Boosting D2D communications in CRNs using POMDP

We propose a novel spectrum prediction (SP) framework that leverages a Partial Observable Markov Decision Process (POMDP) and reinforcement learning (RL) to enhance energy efficiency (EE) in Cognitive Radio (CR) -based Device-to-Device (D2D) communications. Unlike conventional spectrum sensing (SS), which is both energy- and time-consuming, our approach predicts the state of primary user (PU) channels and adopts a hybrid transmission strategy that seamlessly integrates interweave, underlay, and overlay modes. A policy iteration-based RL algorithm enables secondary users (SUs) to dynamically update their transmission strategies based on historical observations and belief states. Furthermore, a dedicated reward function maximizes throughput while minimizing energy consumption under stringent constraints such as data rate, interference thresholds, and PU cooperation requirements. Simulation results demonstrate that our approach outperforms existing ML/RL-based works in EE optimization by $\sim 15.40\text{\%}$ and $\sim 34.86\text{\%}$, respectively.