Resource Selection in Cognitive Networks With Spiking Neural Networks.

This paper explores the feasibility of a spiking neural network-based approach to cognitive networking, that is potentially suitable for low-power neuromorphic chips. We discuss the design of a cognitive network controller (CNC), which can dynamically optimize the selection of resources for recurrent network tasks, based on both its assigned objectives and observations of the actual performance achieved by each resource. We present a coding strategy for the action decisions based on the time-to-fire of spikes, a learning algorithm, and a regulation method to keep synapse strengths within an adequate range. To evaluate the proposed method, we apply the CNC to a challenged network environment using simulation. In this scenario, the CNC requires to optimize the average file transfer time over a multichannel space communication link, which is available only for a time window because of orbital dynamics. Compared to conventional methods, we show that the CNC achieves its objective for a broad range of offered loads. We examine the impact of key system factors that include learning and space protocol parameters. The proposed CNC potentially fosters the development of new cognitive networking applications.