High-order fuzzy cognitive maps and randomized networks for time series and nonlinear dynamical systems

• Ano de defesa: 2023
• Autor(a) principal: Omid Orang
• Orientador(a): Frederico Gadelha Guimarães
• Banca de defesa: Guilherme de Alencar Barreto, Tatiane Nogueira Rios, Hugo Valadares Siqueira, Rodrigo César Pedrosa Silva
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Minas Gerais
• Programa de Pós-Graduação: Programa de Pós-Graduação em Engenharia Elétrica
• Departamento: ENG - DEPARTAMENTO DE ENGENHARIA ELÉTRICA
• País: Brasil
• Palavras-chave em Português: Time series forecasting; Randomized fuzzy cognitive maps; Reservoir computing; Echo state network; Multiple-input multiple-output
• Link de acesso: http://hdl.handle.net/1843/58467

Fuzzy Cognitive Maps (FCMs) have emerged as interpretable Fuzzy Time Series (FTS) methods used in a variety of forecasting applications due to their interesting features. Constructing the structure of FCMs and extracting weighted connections among the concepts compose the crux contribution of the proposed FCM-based approaches in the literature. Despite the success of the proposed methodologies, there are still some gaps and limitations in this domain. To cover some of these challenges, this thesis introduces new forecasting techniques based on FCMs to predict univariate and multivariate time series focusing on both aspects including designing the new architecture and speeding up the training phase. Thus, the main contribution of this thesis is to introduce novel forecasting techniques by merging FTS and FCMs to generate randomized high-order FCM (R-HFCM) as reservoir computing models for the first time in the literature. R-HFCM is a kind of ESN where the reservoir layer consists of a group of sub-reservoirs such that the weights inside each sub-reservoir are randomly chosen according to the ESN weight initialization. The computational experiments demonstrate that R-HFCM outperforms in terms of both accuracy and training speed when compared to the traditional FCMs trained via evolutionary algorithms like genetic algorithm (GA). To fill the absence of Multiple-Input Multiple-Output(MIMO) models, extensions of the univariate R-HFCM method are presented to handle low-dimensional and high-dimensional time series forecasting. It is worth noting that in both MIMO methods, only the output layer is trainable using the time-effective least squares method. The proposed methods obtained promising and competitive results compared with a range variety of deep learning and machine learning methods in terms of accuracy and parsimony.