Applicability of learning curves in performance analysis of Radial basis and Multilayer Perceptron networks for seismic first break picking

Abstract: There are some serious considerations in application of an artificial neural network in a seismic processing stage. In this study the importance of finding optimum minimum number for starting a classification task is discussed. Two structures of Radial basis criteria with Gaussian transfer functions and Multilayer perception architecture is used. Networks shall be carefully turned differently in each structure. Number of layers and number of neurons in each hidden layer are two important parameters to be decided and normally there are no rules finding them precisely. Also the effect of adding seismic random noise on the architecture of MLP & RBF in first break picking is investigated. 1. Introduction training phase of a network There are many types of ANNs, depending on the way the artificial neurons are connected and how the inputs and outputs are linked. Some of them which are widely used in geophysical domain are multi-layer perceptron & radial basis function. In general, a neural network with monotonically increasing activation function is called multilayer perceptron (MLP). Another neural network, which is used widely in oil industry, is Radial Basis Function network (RBF), which is a network with localized activation function such as Gaussian function (Aminzadeh, F. and de Groot P., 2004). RBF network has the same layered structures as the MLP network, except the application of weights and activation function. In this paper a method to find the hidden layer size is described. We used known