Improving reinforcement learning algorithms for dynamic spectrum allocation in cognitive sensor networks

Abstract

Cognitive Radio Networks enable a higher number of users to access the spectrum of frequency simultaneously. This access is possible due to the implementation of dynamic spectrum allocation algorithms. In this context, one of the main algorithms found in the literature is the reinforcement learning based approach called Q-Learning. Although been widely applied, this algorithm does not take into account accurate information about the behavior of users neither the channel propagation conditions. In this sense, we propose three improvements to the dynamic spectrum allocation algorithms based on reinforcement learning for cognitive sensor networks. Simulation results show that all the proposed algorithms allow allocating channels with up to 6dB better quality and 4% higher efficiency than Q-Learning.