Deep Reinforcement Learning for Spectrum Allocation in Cognitive Radio Networks with Realistic Environment Modelling
Seshendranath Balla Venkata, B M Manjula, Mohhamied Husaein Sallaah, Srinivas Samala, Sarva Mangala Praveena
Abstract
Over past few years, Cognitive Radio Networks (CRNs) have emerged as an effective solution for addressing spectrum scarcity by enabling dynamic access to underutilized frequency bands. However, traditional spectrum allocation techniques and earlier Deep Reinforcement Learning (DRL) approaches depend on oversimplified simulation environments, lack realistic channel modelling, and fail to account for sensing inaccuracies and user mobility, leading to limited real-world applicability. Therefore, a Realistic Environment-based DRL Spectrum Allocation (REDRLSA) framework is proposed that integrates Deep Q-Network (DQN) and Quantile Regression DQN (QR-DQN) with an enhanced cognitive radio simulation environment. Initially, spectrum occupancy data is collected from Reference Geolocation Spectrum Database (RGSD) and preprocessed to generate binary channel availability maps. Then environment is augmented with log-normal shadowing, path loss, small-scale fading, probabilistic sensing noise, stochastic primary user activity, and secondary user mobility. Further, agents interact with environment using a Signal-to-Interference-plus-Noise Ratio (SINR)-aware reward function to optimize channel selection while minimizing interference. Experimental results demonstrate that REDRLSA achieves higher interference-avoidance rate of 98 %, improved throughput of 2.02 packets, and robust performance of 96 % under variable channel and traffic conditions, outperforming baseline models trained in simplified environments.