Effect of Training Algorithms and Network Architecture on the Performance of Multi-Band ANN-Based Path Loss Prediction Model

Abstract

The propagation of Electromagnetic waves signal in terrestrial frequency bands in a build-up environment is affected by many factors, leading to signal degradation, diffraction, reflection, scattering, among others. Furthermore, the physical layer interface is one of the most critical factors needed to be carefully analyzed for optimum design of wireless systems. In view of these, several channel models were proposed to optimally predict how radio waves behaves in a typical built-up and complex environments. In the case of Nigeria, when these propagation models are considered at disparate environment, a many of them are susceptible to tremendous prediction error. Hence, there is a need to develop a model suitable for such an environment to minimize errors. This paper used Scale Conjugate and Levenberg-Marquardt algorithms to develop a multi-frequency bands ANN-based path loss prediction model. Furthermore, the paper investigated the effect of ANN system parameters on the model's performance. Findings revealed that Standard Deviation Error (SDE) and the Correlation Coefficient (R) depend on the model's network architecture. In addition, the Levenberg-Marquardt algorithm fits the network with complex structures compared to the scale conjugate algorithm. It was further discovered that increment of the number of hidden neurons, ordinarily, does not, in the same way, means an increase in the performance of the model.