Optimizing Multilayer Perceptrons to Approximate Nonlinear Quaternion Functions

In this contribution, a novel approach to optimize multilayer perceptron artificial neural networks (MLP-ANN) devoted to approximate quaternion valued functions is presented. The approach is based on the definition of proper auxiliary networks devoted to predict the trends of the main network weights during the learning phase, thus reducing the number of epochs needed to reach a suitable abstraction level. The approach is specifically designed for MLP-ANN devoted to the approximation of complex valued functions, characterized by more than one imaginary part, a framework typical of application-oriented problems.