Weight Training for a Multilayer Perceptron: A Comparison Study

This paper presents a critical comparison of four optimisation algorithms for training a multilayer perceptron. The chosen algorithms are backpropagation, simulated annealing, genetic algorithms and Bayesian learning. We use a petroleum reservoir data set to compare the performance of these algorithms. The data set is randomly splitted into a training set and a test set. Error bounds are generated for all the test data. We use various statistics as performance indicators. The study shows that simulated annealing is the best algorithm for fast and efficient learning of the data set.