Neuro-Cognitive Radios for Dynamic Spectrum Access

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 for the first time a neuromorphic-based learning architecture to address a challenging decentralized DSA scenario where multiple source-destination pairs share a limited number of spectrum bands. Sources, called Neuro-Cognitive Radios (NCR), run spiking neural network (SNN)-based reinforcement learning (RL) architectures that adapt their transmission strategy over time, in a decentralized manner, aiming to maximize their own throughput while striving for network-wide fairness. We evaluate NCR in several network settings and compare with an equivalent Deep Q-Network (DQN) architecture that uses multi-layer perceptrons as opposed to spiking neurons. Our simulation results show that NCR outperforms DQN in terms of fairness while keeping a competitive throughput level, thereby demonstrating the potential of neuromorphic computing to address DSA problems.