Improved classification performance of q-Gaussian meta-cognitive RBF classifier

In this paper, a new meta-cognitive RBF neural network classifier that uses a q-Gaussian activation function is presented. The q-Gaussian activation function has the capability to extend or contract the shape/response of the radial basis activation function, based on the value of the parameter q. This property is used to avoid a sharp fall in the response in the tail region, particularly when the feature space has large dimensions. In this paper, additionally the projection based learning algorithm developed earlier has also been improved by including additional criteria for neuron addition and parameter update. Using both binary and multiclass benchmark datasets with various level of class imbalance, from the UCI repository, the performance of the new classifier (referred to as Extended McRBFN) has been evaluated and compared with other existing Projection based learning-metacognitive RBF network (PBL-McRBFN), Extreme Learning Machine (ELM) and Support Vector Machine (SVM) classifiers. The results indicate that EMcRBFN achieves better classification performance compared to other classifiers for most of the datasets and at least similar performance for the remaining datasets.