Evaluating classification performances of single-layer perceptron with a Choquet fuzzy integral-based neuron

The sigmoid function is usually used as the activation function for a well-known classification method, namely the single-layer perceptron. In the function, a weighted sum, in which the additivity among individual variables is assumed, is performed. However, it is known that an assumption of additivity may not be reasonable, since the input variables are not always independent of each other. This paper thus employs a Choquet fuzzy integral-based neuron as an output neuron of the single-layer perceptron. Moreover, the connection weights can be interpreted as fuzzy measure values or degrees of importance of the respective attributes. The connection weights are determined by the genetic algorithms in which the maximization of the training classification performance and the minimization of the errors between the actual and desired outputs of individual training patterns are taken into account. The experimental results further demonstrate that the classification results of the single-layer perceptron with a Choquet fuzzy integral-based neuron are comparable to those of the traditional single-layer perceptron and the other fuzzy classification methods.