Multi-Layer Perceptron and a Radial Basis Function Architectural Network Models for Signal Propagation Power Loss Prediction – Pros and Cons

—This paper designs and applies a multi-layer perceptron neural network model and a radial basis function neural network model to predict signal power loss. It compares their advantages and disadvantages in predicting signal propagation environments using measurement data from a long-term evolution line-of-sight environment. Thus, a major focus is on their performance in prediction accuracy and efficiency. The details of the applied models, their architectural structures, and their geometries are analyzed, as these greatly influence their performance. The models were trained using data collected from a realistic, complex line-of-sight, diverse environment. The simulations allowed the trained multi-layer perceptron and the radial basis function network models to quickly simulate new geometries of significant complexity within the dataset. Essentially, this addresses the trade-off between efficiency and accuracy in ANN propagation models. The regularization technique and early stopping training method were used to ensure proper network generalization, with neuron counts in steps of ten in the hidden layers during training. The number of neurons in the hidden layers of both models impacts their predictive capabilities. Fewer neurons in the multi-layer perceptron led to underfitting, whereas too many causes overfitting. A moderate number of 40 neurons in the hidden layer shows good generalization. The radial basis function network, however, offers better and more efficient predictions of complex systems compared to the multi-layer perceptron, as its accuracy improves with more hidden layer neurons. At 70 neurons, the correlation coefficient was nearly +1. It also takes less training time for the radial basis function network to predict signal power loss than the multi-layer perceptron.