Recurrent Flash Reinforcement Learning for Dynamic Spectrum Access and Power Control

The growing number of Internet of Things (IoT) devices has increased demand for wireless spectrum, exacerbating the spectrum scarcity issue. While dynamic spectrum access (DSA) has emerged as a promising solution, most existing works focus on channel selection, neglecting other control factors. To address this, we develop a deep reinforcement learning (DRL) environment that involves both spectrum access and power control, and propose a dual-head output DRL structure that efficiently handles this mixed action space in a unified model. Moreover, we represent Recurrent Flash (ReF) framework, which integrates linear attention with a recurrent mechanism to enhance the temporal memory abilities of agents. Utilizing ReF as representation learning layers of DRL approaches enables a more precise comprehension of environment dynamics and promotes better-informed decision-making. The combination of the dual-head output and ReF, referred to as ReF-RL, is rigorously evaluated through experiments. Results demonstrate that ReF-RL significantly improves communication quality and device longevity, thereby outperforming both conventional communication and existing DRL algorithms.