Uninorm based regularized fuzzy neural networks

This paper proposes a training algorithm for fuzzy neural networks that can generate consistent and accurate models while adding some level of interpretation to applied problems. Learning is achieved through extreme learning machine concepts, allowing the adjustment of parameters during the training phase using a fast and straightforward approach. The use of the regularization in the inner layers of the model will enable it to be more precise and selfish since a reduced set of fuzzy rules can be extracted from the final result of the network. The proposed approach was evaluated through pattern classification problems using real datasets of large and small sizes. The achieved results were compared to the results obtained using another state of the art classifiers. Statistical analysis of the results suggests the proposed approach as a promising alternative to performing classification with some level of model interpretability.