Machine learning algorithms to solve statistical problems

• Autor(a) principal: Souza, Fábio Hemerson Araújo de
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Institucional da Universidade Federal do Ceará (UFC)
• Texto Completo: http://www.repositorio.ufc.br/handle/riufc/56651
• topic: Inteligência artificial; Aprendizado de máquina; Algoritmos; PIPE; Redes neurais; Artificial inteligence; Machine learning; Algorithms; Neural Network

The field of artificial intelligence before its popularity, begin to be spread around the world as a computational tools from a distant future. In 1950, Artificial Neural Network (ANN) begin to be developed and all compute intelligence algorithms follow the same way , making that future be more closer now a days. Currently, machine learning, an AI branch, is used in many different fields and processes such as marketing and sales, business intelligence, research and development, supply chain, financial stocks, human resources, healthcare, etc. The maturation of AI field carries itself a probabilistic bases that can be used to solve some problems in statistics. In this work we make use of machine learning techniques and algorithms to solve two proposed statistical problems. In chapter 1, the issue is to find an approximation to normal cumulative distribution function. This expression needs to be mathematically and computationally simpler than other approximations founded in statistics lectures and papers and one possible use of this expression is in introductory statistics classrooms. Chapter 2 we address an identifiability distribution problem using machine learning algorithm and a framework for mathematical computation called Tensorflow and an abstraction library for deep learning routines called Keras, both of them written in Python. The main goal here is construct a structure that can be able to capture features from a sample provided by the user and classify the parent distribution of this sample. The results were promising with a accuracy greater then 95% for each distribution used for examples.