Spectrum Inference in Cognitive Radio Networks with Machine Learning

Abstract

Wireless radio spectrum is a limited resource. Increasing demand for more spectrum bands has led to the notion of its efficient and intelligent utilization. Cognitive radio technology is the front runner in dynamic spectrum access. Basic spectrum management tasks of sensing, mobility, sharing and decision have been improved by using machine learning techniques. Real time sensing and related operations thereafter involve considerable time delays leading to decreased throughput. Spectrum Inference has emerged as an effective solution to this problem. In this work we have analyzed machine learning based spectrum inference techniques for real world dataset. Spectrum band occupancy prediction has been formulated as a regression problem. Three regression based approaches namely linear regression interactions, SVM based regression and decision tree regression have been evaluated. It has been observed that fine tree regression gives the best performance. To optimize the performance in terms of prediction speed and accuracy we have investigated the use of improved and increased number of features. With addition of a single additional feature the prediction speed has increased by 4.73 times and prediction accuracy by 3%. However the training time has increased by 1.24 times.