A Machine Learning Based Scheme for Dynamic Spectrum Access

Abstract

In this paper, we present a machine learning (ML) based dynamic spectrum access (DSA) scheme which can be used in a system in which the primary user (PU) spectrum occupancy can be represented as a sequence of busy (on) and idle (off) periods. We use real world data collected from Long Term Evolution (LTE) systems at two locations for our study. We experiment with different feed forward artificial neural network (ANN) architectures to choose from for our DSA scheme. A simple perceptron based ANN architecture was determined to provide good performance. We compare performance of our ML based DSA scheme with a traditional DSA scheme based on analytical model that uses survival analysis. Our results show that our ML based scheme outperforms the survival analysis based scheme in terms of utilization of idle periods. In terms of probability of interference to the PU, our scheme is better in some configurations and slightly worse in some other configurations.