Methodological comparison of machine learning techniques to improve adaptability and reduce the handoff rate in cognitive radios

Abstract

A large number of devices connected in the Industrial, Scientific & Medical (ISM) bands, in large urban centers, makes interference between them inevitable. Analyzing this, the Federal Communications Commission (FCC) assembled in 2002 a team to improve the policy of separation of the electromagnetic spectrum: the Spectrum Police Task Force (SPTF). Then the Cognitive Radio (CR) were proposed as a solution, allowing secondary users to connect on frequencies reserved for primary users, with low interference between them. One of the functions of CR is the automatic selection of channels in the electromagnetic spectrum. Several algorithms have been proposed to predict which will be the next channel, but few are concerned with the geographic adaptability of the model and the number of handoffs that the radio does. In this work, we propose a methodological comparison between the CRF and Q-Learning, to analyze the adaptability of these two distinct geographic local techniques, maintaining precision and a low handoff rate. Testing on Wi-Fi frequencies the CRF shows more adaptive, due its temporal windows, on average with 97.6% adaptability versus 95.68% for Q-Learning. In addition, a CRF handoff rate remained below 0.1% in all locations and frequencies, against an average of 5.8% for Q-Learning.