Entropy Based Exploration in Cognitive Radio Networks using Deep Reinforcement Learning for Dynamic Spectrum Access

This paper details the practical design of a Cognitive Radio network which uses multi-agent Deep Reinforcement Learning for dynamic spectrum access. Each network node evaluates a neural network model to determine when it can transmit and on what frequency channel. The models are trained offline in simulation to mitigate slow online training time. Furthermore, we propose the use of entropy-based-exploration to dynamically determine when more training is required in the wireless network. Unlike previous work that has only considered similar techniques in theory and simulation, we present over-the-air measurement results for the throughput and channel utilization collected in a large-scale software-defined radio testbed.