Use of an artificial neural network-based metamodel to reduce the computational cost in a ray-tracing prediction model

Abstract

The purpose of this article is based on analyzing the use of RTQ3D ("quasi-3D" ray tracing technique) to produce the value of the initial electromagnetic fields or fitness for a hundred and sixty receivers according to the possible positions of two antennas to be distributed in a closed environment. The problem variables consist of the values of the magnetic fields for one hundred and sixty receptors depending on the positions of the antennas to the base stations, which serve as input data for the algorithm to the RMLP (Artificial Neural Network, multilayer perceptron with Real backpropagation learning algorithm). The values of the magnetic fields associated with the positions of the antennas are the values to be learned by the network, the teacher of RMLP. This study aims to develop efficient techniques for optimization of electromagnetic problems. We use the PSO (Particle Swarm Optimization) algorithm associated with a metamodel based on an ANN (Artificial Neural Network). Specifically, we use the MLP (Multilayer Perceptron) with the backpropagation algorithm in order to evaluate objective functions in an efficient way. The ANN will be used to assist the technique of "quasi 3D" ray-tracing in order to reduce the high computational cost of this technique in PSO optimization.