Comparison Between Different Architectures of Multilayer Perceptron Networks for Blocking Rate Prediction in Mobile Phone Networks

Blocking in mobile phone networks is a problem that consists of the refusal of the connection between a telephone device and a cell responsible for emitting the signal. The occurrence of blocking can indicate that a cell is close to congestion, leading to financial losses for telephone companies. This work developed three prediction systems using Multilayer Perceptron neural networks. Each system was modeled following different strategies: Direct, Recursive, and Direct Recursive, respectively. The training and test of the networks were carried out by using real data containing the history of blocking rates from a network of cells. The development stages consisted of analyzing the performance of each prediction system, varying the number of neurons in the hidden layers and the number of predicted steps from 1 (corresponding to 15 minutes ahead) to 20 (corresponding to 5 hours ahead). The system based on the Recursive strategy presented the lowest performance making predictions of short (15 minutes) and long (5 hours) terms with RMSE (Root Mean Squared Error) of approximately 13% and 40%, respectively, with a confidence interval between 27% and 29% considering all predictions. The systems based on the Direct and Direct Recursive strategies presented similar results, making predictions of short and long terms with RMSE of approximately 12% and 31%, respectively, with confidence intervals between 21% and 23% considering all predictions. Although the Direct and Direct Recursive systems obtained the lowest RMSE, the Direct Recursive is more advantageous as it requires fewer MLP networks. Consequently, it has simpler training and a lower computational cost.