Usually neuro-fuzzy networks are trained by using mean squared error (MSE) as cost function. This approach may lead to some problems when the network is used to classify patterns. In this paper, we propose a new empirical risk functional that is particularly suitable for classification tasks, when neuro-fuzzy learning is based on a gradient descent strategy. This functional has the properties of being a good misclassification-rate estimator and, at the same time, of being differentiable, so it can be used in the error back-propagation algorithm. The new functional has been compared with standard MSE on the well-known Iris data benchmark. Experimental results have shown that the use of this new cost function leads to very fast convergence with high classification rates.
Scheda prodotto non validato
Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo
|Titolo:||A new empirical risk functional for a neuro-fuzzy classifier|
|Data di pubblicazione:||2000|
|Appare nelle tipologie:||4.1 Contributo in Atti di convegno|