Deep Reinforcement Learning for Dynamic Spectrum Allocation in Cognitive Radio Network

Abstract

The wildest boom of wireless devices and the shift to 5G/6G ecosystems contributed to the lack of the spectrum, making the old traditional methods of static allocation less and less efficient. cognitive radio networks provide an alternative with dynamic nature, the current solutions tend to fail because of the sophisticated nature of imperfect channel state information, large-dimensional state space and the rapid mobility of users. This model presents an Attention-Augmented Multi-agent Deep Reinforcement Learning model, which is used to maximize autonomous spectrum sharing by using spatial-temporal awareness. The architecture is based on convolutional neural networks in mapping spatial interference and long short-term memory layers in temporal mobility tracking with a multi-head self-attention mechanism to coordinate interference management between secondary users. To obtain accurate resource mapping, layers of Sinkhorn are incorporated to be bi-stochastic. Simulation shows that the spectral efficiency is 22.14% higher and the collision rate is also 6.82 times lower with the 3GPP channel models than with regular deep Q-networks. The system can be 85.36% efficient even in the presence of serious channel state errors, which is a strong solution to ensure trustworthy ultra-dense urban connectivity.