Investigation and Comparison of Generalization Ability of Multi-Layer Perceptron and Radial Basis Function Artificial Neural Networks for Signal Power Loss Prediction

摘要

This research work investigates, compares and presents three different approaches to avoid poor generalization and to overcome the tendencies of over-fitting in Radial Basis Function (RBF) and Multi-Layer Perceptron (MLP) artificial neural networks, in order to enhance the prediction accuracy of signal power loss during electromagnetic signal propagation in metropolitan area using measured data from a Long-Term Evolution (LTE) network. These approaches are a variation of hidden layer neurons, early stopping and Bayesian Regularization techniques. In the variation of the hidden layer neurons in MLP network, an excellent prediction has been recorded with 40 neurons while an increase in the number of neurons leads to poor network generalization. The network shows the capability of modeling a moderate size propagation network and, for more complex networks, intermediary layers or network modifications are required. For RBF network, the generalization ability of the network increases as the network gets more complex with 70 neurons in the hidden layer giving the best prediction. Training RBF network using early stopping approach gives a better prediction with less errors compared to neuron variation in the hidden layer and Bayesian Regularization in MLP network. However, in RBF network training, there is no difference in the errors obtained when early stopping approach has been used compared to Bayesian Regularization approach. It models appropriately a complex network without signs of over-fitting, and because of its fixed three-layer architecture, there is no poor generalization resulting from architectural complexity which Bayesian regularization approach tackles.