- author = "Glender {Bras de Medeiros}",
- title = "Abordagens evolucionarias para aprendizado autonomo em redes neuro-fuzzy preditivas",
- school = "Centro Federal de Educacao Tecnologica de Minas Gerais",
- year = "2025",
- address = "Belo Horizonte, Brazil",
- month = "26 " # aug,
- keywords = "genetic algorithms, genetic programming, CEFET-MG, Neuro-fuzzy networks, autonomous learning, evolutionary fuzzy systems, evolutionary computation",
-
URL = "
https://sig.cefetmg.br/sigaa/public/programa/defesas.jsf?lc=en_US&id=308#7",
-
URL = "https://sig.cefetmg.br/sigaa/verArquivo?idArquivo=6110333&key=15bfe5322b9ac0c22fbfff979b35d96a",
- size = "108 pages",
- abstract = "This work proposes new evolutionary approaches for autonomous learning in neuro-fuzzy networks. Three new neuro-fuzzy models are presented: one linguistic (FALCON-GA) and two functional (EvoNFuzz and DGP-NF). In FALCON-GA, a genetic algorithm is used for rule extraction, and a gradient-based method is used for parameter adjustment. The EvoNFuzz and DGP-NF models use multi-gene genetic programming to model functional rules. They also present a structure closer to neural networks, incorporating synaptic weights into the rule consequents. For parameter adjustment, EvoNFuzz uses a gradient-based method, while DGP-NF implements an incremental learning method based on weighted least squares. Furthermore, DGP-NF incorporates mechanisms to improve the search space exploration, such as dynamic population and population diversity control, and a pruning operator based on gene diversity analysis to eliminate redundant rules. The proposed mechanisms are based on phenotypic diversity, aiming to improve the quality of solutions generated by genetic programming and the responses provided by the final model. Experimental results on 21 nonlinear regression datasets suggest that the proposed approaches are promising and that the proposed models are efficient for this type of problem, with DGP-NF standing out as the best performer on 16 datasets and competitive on the remaining datasets. The results also suggest that the incremental learning method outperforms the gradient method for parameter tuning in evolutionary neuro-fuzzy networks and that functional networks are more efficient than linguistic networks for regression problems. The use of dynamic population and diversity control mechanisms proved effective in avoiding premature convergence and improving model accuracy. Furthermore, phenotypic diversity-based pruning has demonstrated potential as a simple and effective strategy for pruning multi-gene individuals and eliminating redundant rules.",
-
notes = "In Portuguese
Supervisor: Alisson Marques da Silva",
