Risk-aware dynamic spectrum allocation using learning-augmented RIS control for cognitive radio networks

Abstract

The proposed RADIANT-CRN framework introduces a risk-aware dynamic spectrum allocation approach for reconfigurable intelligent surface (RIS)-enabled cognitive radio networks while maintaining reliable protection for primary users (PUs). It incorporates our previously developed bidirectional long short-term memory and adaptive manta-ray foraging optimization (BiLSTM-AMRFO) spectrum prediction and deep channel estimation models into a bilevel optimization framework, where semidefinite relaxation (SDR) adjusts RIS phase shifts, geometric programming (GP) allocates transmit power, and entropy-regularized assignment performs channel selection. A primal-dual actor–critic framework coordinates these modules. On a 120 MHz testbed with 256-element RIS, RADIANT-CRN gets a sum rate of 846±18 Mbps with a 0.7±0.2% PU-violation and 99.3% chance-constraint coverage. This is about 25% higher than a greedy non-risk baseline and about 43% higher than a no-RIS optimizer. It also lowers interference CVaR_0.95 from 9.1 mW to 2.4 mW. These results demonstrate that RADIANT-CRN is the first framework to enforce both chance and CVaR guarantees in RIS-assisted CRNs, achieving high spectral efficiency with statistically certifiable PU protection that aligns with Federal Communications Commission interference requirements. The framework is validated on a prototype SDR testbed (Rician fading, K=6 dB, bursty PU activity); implementation uses Python 3.11/PyTorch 2.x with convex optimization python (CVXPy/MOSEK), and synthetic testbed traces are available upon request.