Artificial Neural Networks (ANN) have been widely used in science and engineering problems. They attempt to model the ability of biological nervous systems to recognize patterns and objects. ANN basic architecture consists of networks of primitive functions capable of receiving multiple weighted inputs that are evaluated in terms of their success at discriminating the classes in Τa. Different types of primitive functions and network configurations result in varying models (Hastie et al., 2009; Rojas, 1996). During training network connection weights are adjusted if the separation of inputs and predefined classes incurs an error. Convergence proceeds until the reduction in error between iterations reaches a decay threshold (Kotsiantis, 2007; Rojas, 1996). We use feed-forward networks with a single hidden layer of nodes, a so called Multi-Layer Perceptron (MLP) (Venables and Ripley, 2002), and select one of two possible parameters: size, the number nodes in the hidden layer.
