Learning and prediction using radio and non-radio attributes in wireless systems

Wireless technology and connectivity are spreading rapidly around the globe. The advancement of machine learning tools plays a major role in pushing the limits of wireless communication applications. With the volume, variety and velocity of rich datasets available, we are able to accurately model and predict different aspects of wireless systems which can help with efficient utilization of available resources. We aim to harness radio (such as channel capacity and received signal constellation) and non-radio attributes (such as weather, busy period data from an open-source API) by applying machine learning algorithms to improve the overall wireless system. In this dissertation, we start by working on a small scale network. We present the problem of jointly predicting modulation and the number of transmit antennas for a non-cooperative system or a dynamically operating cognitive system. We use a discrete-wavelet transform on the received complex samples to separate the different modulation classes. We then use the k-nearest neighbor approach coupled with k-means clustering to further classify the signals, utilizing the symmetry and relative distances of the constellation points. The performance of the classifiers at different stages of the algorithm demonstrates a high accuracy where the packets can be decoded. This is followed by predicting channel transition and the state of the environment to improve the existing transmit beamforming procedure by changing sounding times. The experimental results using Software Defined Radios (SDRs) on the ORBIT testbed show more than 96% classification success of channel transition and a significant difference in error vector values for different types of channel variations. We then progress to working on a larger scale, by employing deep learning techniques on a cellular network in a dense urban area using statistically significant non-radio features. We focus on using non-radio attributes and their impact in predicting cellular traffic. Prediction of user traffic in cellular networks is one of the promising ways to improve resource utilization among base stations. We consider traffic from neighboring cells and other non-cellular-traffic related attributes such as weather, busy period data from open-source API as features to augment the cellular traffic data and improve prediction. Specifically, we augment cellular traffic data from the City of Milan and its surroundings and we perform two types of analyses: (i) a one-step prediction or a point-by-point forecast of traffic and (ii) a trend analysis which is the forecast of traffic over an extended period. We compare the results with existing statistical methods such as auto-regression integrated moving averages (ARIMA) and exponential smoothing and observe gains in the trend analysis by providing the augmented data, whereas the one-step prediction is not much impacted.Ph.D.Includes bibliographical reference