Article Preview
Top1.0 Introduction
Telecommunication is an established cutting-edge technology that permits two parties to communicate employing voice and data signals (Rappaport 2002). The evolution, deployment, and application of various cellular radio frequency (RF) based telecommunications systems has orchestrated rapid development in every aspect of human endeavour (Imoize et al. 2021). Starting from the first generation (1G) brand that came into existence in the early 80s to the ubiquitous fourth generation (4G) LTE (Imoize et al. 2019) (Imoize and Oseni 2019), the recently commercialized fifth-generation (5G) (Gupta and Jha 2015), and the envisioned sixth-generation (6G) wireless systems (Dang et al. 2020), the telecommunication industry is progressive globally. The 4G and 5G systems are data communication-centric (Shynu and Al-Turjman 2021). 4G LTE can provide robust data throughput rates in open terrains (Imoize and Adegbite 2018). However, in built-up terrains such as dense urban cities, where all forms of interference and multipath fading impacts are high, 4G LTE may experience low and fluctuating throughput data rates. Hence, there is a need to regularly conduct a measurement-based prognostic examination of User Equipment (UE) throughput rates (Ughegbe, Adelabu, and Imoize 2021). These rates help the RF engineers make the necessary optimization decisions, detect and mitigate interference, including other anomalies that could negatively impact 4G LTE network performance (Huang et al. 2013).
Prognostic algorithms are central to a detailed examination of user data throughput rates (Estevez, Orchard, and Kailas 2013). In recent years, Machine Learning (ML) models and their prognostic algorithms have been deployed for diverse applications comprising data scrutinizing, learning, pattern close-fitting, knowledge extracting, knowledge deducing, and others (Alvarez, Louveaux, and Wehenkel 2017), (Song, Ristenpart, and Shmatikov 2017). There exist numerous ML-based prognostic regression models in the literature (Bui and Turner 2014; Cao and Fleet 2014; Chen and Ren 2009; Chen and Wang 2018; Dervilis et al. 2016; Gu and Hu 2012; Skilling 2006; Su, Peng, and Hu 2017; Vanhatalo, Pietiläinen, and Vehtari 2010; Wan and Ren 2015). One of the most promising techniques is the Gaussian Process Regression (GPR) (Rasmussen 2004) (Chalupka, Williams, and Murray 2013) (Wilson and Adams 2013). The GPR is a non-parametric Bayesian modeling technique with a Gaussian probabilistic structure. Some of the critical advantages of the GPR include (Bui and Turner 2014; Cao and Fleet 2014; Chen and Ren 2009; Chen and Wang 2018; Gu and Hu 2012; Vanhatalo et al. 2010): (i) relatively simple parameterization and implementation structure (ii) proficiency in dealing with stochastic processes of varied intricacies and complexities (iii) proficiency in adaptive learning of noisy and non-noisy data (iv) adeptness in handling uncertainties in datasets (v) expert knowledge incorporation (vi) flexible input data probability distributions, (vii) capacity to integrate prior knowledge (viii) precise stationary and non-stationary fitting of input-output data capability (ix) ability to estimate posterior degradation.