Performance Comparison of Non-Linear Median Filter Built on MLP-ANN and Conventional MLP-ANN: Using Improved Dataset Training in Micro-Cell Environment

Abstract

This research work explores the Levenberg- Marquardt training algorithm used for Artificial Neural Network (ANN) optimization during training and the Bayesian Regularization algorithm for the enhanced generalized trained network in training a designed non-linear vector median filter built on Multi-Layer Perceptron (MLP) ANN called model-1 and a conventional MLP ANN called model-2. The model-1 employed in the design helps in dataset de-noising to ensure the removal of unwanted signals for the improved training dataset. An early stopping method in the ratio of 80:10:10 for training, testing, and validation to overcome the problem of over-fitting during network training was employed. First-order statistical indices, the standard deviation, root mean squared error, mean absolute error, and correlation coefficient were adopted for network training analysis and comparative analysis of the designed model-1 and model-2, respectively. Two locations, Line-of-sight (location-1) and non-Line-of-Sight (location-2), were considered where the dataset was captured. The training results from the two locations for the two models demonstrated improved prediction of signal power loss using model-1 in comparison to model-2. For instance, the correlation coefficient, which shows the strength of the predicted value to the measured values (closer to 1) establishing a strong connection, gives 0.990 and 0.995 using model-1 for location-1, training with Lavenberg-Marquardt and Bayesian Regularization algorithm, respectively and 0.965 and 0.980 for model-2 using the same algorithms. It is seen that the Bayesian regularization algorithm, which optimizes the network in accordance with the Levenberg- Marquardt algorithm, gave better prediction results. The same sequence of improved perditions using designed model-1 in comparison to model-2 were seen with training results in location-2 while also adopting other employed 1st order statistical indices.