Neuro-Cognitive Radios for Dynamic Spectrum Access

ICLR 2026 Conference Submission19898 Authors

Keywords: reinforcement learning, spiking neural networks, dynamic spectrum access, wireless networks, cognitive radio

Abstract: Neuromorphic computing is an emerging brain-inspired information processing paradigm that is well-suited for energy-efficient, real-time, and adaptive applications such as Dynamic Spectrum Access (DSA). In this paper, we develop and present the Neuro-Cognitive Radio (NCR) framework$-$a neuromorphic-based learning architecture to address challenging decentralized DSA scenarios where multiple source-destination pairs share a limited number of spectrum bands. NCR combines spiking neural networks (SNNs) and reinforcement learning (RL) architectures that adapt their transmission strategy over time to maximize overall throughput and fairness. We evaluate NCR in several network settings (with and without jamming) and compare with an equivalent Deep Q-Network (DQN) architecture that uses traditional multi-layer perceptron (MLP). We show that NCR achieves higher fairness than DQN, while keeping a competitive throughput level. This work constitutes a promising initial step toward neuromorphic-based frameworks for solving DSA problems.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 19898