V3I5P58

Reinforcement Learning-Based Spectrum Management in Cognitive Radio Systems 

Abdul-Azeez Dauda^1*, Muhammad Babadoko Muhammad²

Abstract

Due to the recent widespread growth of fifth-generation (5G) wireless networks and upcoming sixth-generation (6G) ones, scarcity of spectrum has been more relevant than ever, since traditional frequency allocation frameworks are becoming less applicable. This paper proposes a comparative reinforcement learning (RL) framework for dynamic spectrum access (DSA) in cognitive radio (CR) systems, filling an important gap identified in previous work: The inability of single-agent RL models to preserve performance stability under non-stationary channel dynamics and dense secondary user (SU) deployments. We tested on OpenAI Gym compatible CR simulation environments and compared the performance of Q-Learning, Deep-Q Network (DQN), and Multi Agent Reinforcement Learning (MARL) with 10000 training episodes with different PU traffic loads. DQN surpassed random access baselines by 34.7% in spectrum utilization, and reduced PU interference by 41.2%, while MARL achieved 28.3% lower inter_SUs collision rates against non-MARL methods in high-density scenarios. These findings position RL-empowered spectrum management as a disruptive paradigm shift for adaptive CR systems, with immediate relevance to 6G edge intelligence and real-time spectrum governance.

Keywords:

cognitive radio, deep Q-network, dynamic spectrum access, multi-agent reinforcement learning, spectrum management