How to Learn Fading from Data

• Grant number: 19/04258-9
• Support type: Scholarships abroad - Research
• Effective date (Start): May 01, 2019
• Effective date (End): September 30, 2019
• Field of knowledge: Engineering - Electrical Engineering - Telecommunications
• Principal researcher: José Cândido Silveira Santos Filho
• Grantee: José Cândido Silveira Santos Filho
• Host: Flavio du Pin Calmon
• Home Institution: Faculdade de Engenharia Elétrica e de Computação (FEEC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Research place: Harvard University, United States

Abstract

Relying on physical arguments, researchers have proposed different probabilistic fading models to describe the random fluctuations that impair the amplitude and phase of the wireless communication channel. Also, for each of these models, they have employed classical approaches to estimate the distribution parameters from field measurements, including the ubiquitous moment-matching and maximum-likelihood methods. This is how fading channels have been mostly learned from data so far. In this project, we revisit this problem from a more pragmatic, yet more fundamental point of view. Regardless of the functional form of the true fading data distribution, which is in practice unknown, what is the unified and essential aspect of fading that dominates system performance? What is a simple set of parameters to describe this aspect appropriately? What is a suitable model to learn these parameters from fading data? Which cost function should be assigned to the learning process? Which minimization algorithm would fit best? We aim to answer all of these questions by building on a mix of domain-specific wireless communications insights and data-driven machine learning techniques.