Bayesian fuzzy clustering neural network for regression problems

Smart models are responsible for solving complex problems within the routines of people and companies. The complexity of their structures can be defined by the methodologies used in the definition of parameters or architecture of their models. In three-layer fuzzy neural networks, the number of neurons is determined by fuzzification techniques capable of creating elements based on the data submitted to the training of the intelligent structures. The choice of the fuzzification method can make the model architecture complex and with many parameters. To create a more compact fuzzy neural network architecture, this paper presents a new fuzzy approach for fuzzy neural networks based on Bayesian clustering. This technique expands the training capabilities of the fuzzy c-means algorithm by adding to its methods the nature of the particle filter inference technique to estimate the model parameters including the number of clusters, making the approach more straightforward and less dependent on settings for the construction of the neural network fuzzy. The other layers of the model use fuzzy logical neurons capable of creating IF / THEN rules and use the extreme learning machine to update parameters. The proposed new model was submitted to linear and nonlinear regression databases to confirm that the proposed model can act as a universal approximation of functions. The results presented were promising about the ability to identify patterns within bases in complex problems.