A Logarithmic Neural Network Architecture for Pra Approximation

A neural network based risk monitor was designed to emulate the results of a Nuclear Power Plant probability risk assessment. Although multilayer feedforward neural networks with sigmoidal activation functions have been termed universal function approximators, this approximation may require an inordinate number of hidden nodes and is only accurate over a finite interval. These short comings are due to the standard multi-layer perceptron's (MLP) architecture not being well suited to non-linear function approximation over a large interval. A new architecture incorporating a logarithmic hidden layer proves to be superior to the standard MLP in non-linear function approximation. This architecture uses a percentage error objective function and a hybrid simulated annealing and gradient descent training algorithm. Non-linear function approximation examples are uses to show the increased accuracy of this new architecture over both the standard MLP and the logarithmically transformed MLP. The risk monitor was not fully developed due to the unavailability of the necessary amount of training data required to train the network. However, the lessons learned from this project are valuable to other researchers performing large scale implementations of neural networks.