Effect of Learning Rate on GRNN and MLP for the Prediction of Signal Power Loss in Microcell Sub-Urban Environment
Virginia Chika Ebhota, Joseph Isabona, Viranjay Mohan Srivastava · International Journal on Communications Antenna and Propagation (IRECAP) · 2019
This research work analyses and presents the effect of learning rate in two Artificial Neural Network (ANN) models: (i) Generalized Regression Neural Network (GRNN), and (ii) Multi-Layer Perceptron (MLP) neural network, used for the prediction of signal power loss in a micro-cell environment. Data has been collected from two separate base stations with Long-Term Evolution (LTE) network are used as inputs for training the models. The GRNN and the MLP neural network are trained using Bayesian Regularization training function and 40 neurons in the hidden layer. The data from the first Base Station (BS1) has been used to investigate the effect of 0.5 to 14 spread factors in GRNN and the effect of utilizing different combination of transfer functions in the hidden and output layers of the MLP model during network training. The data from the second Base Station (BS2) has been used in evaluating the performance of the two models considering their prediction accuracy with different variation of learning rate from 0.2% to 1.4%. Performance metrics such as the Root Mean Square Error, the Mean Absolute Error, the Correlation Coefficient and the Standard Deviation are used for analyzing the results. The results using data from BS1 shows a spread factor of 0.5 with best prediction ability and the least error in GRNN. Small spread factor generalizes a GRNN model excellently even with a high learning rate. Also, hyperbolic tangent and logistic transfer functions in the hidden and output layer of MLP network give the best prediction accuracy for the problem under consideration. The results with data from BS2 show GRNN with the least prediction error at a learning rate of 0.2% with 85% of trained and tested data in the error histogram with 20 bins. The MLP network predicted accurately at a learning rate of 1.0 % with 70 % of the measured data trained and tested in the error histogram with 20 bins.