Aprendizado de máquina aplicado à gerência de redes sem fio

Abstract

With the rapid increase of mobile users, the pursuit of resource management in wireless networks has become an important topic. Because of this, studies on predicting mobility and volume of users connected to a network are relevant, especially considering recent advances in machine learning techniques. To this end, understanding the mobility of users, in order to predict the next access point for connection, generates benefits in the scope of resource management such as the search for the much desired transparent handoff. Thus, predicting the number of users, and by volume, we mean the number of users connected to the network, brings great opportunities such as managing the available network bandwidth and avoiding the expense of unnecessarily connected access points. Therefore, this work uses two sets of real data originating from two university fields (UFJF and KTH), to train predictive models based on historical and recent data. In this work, the performance of these models was analyzed for mobility and number of users prediction approaches using the following algorithms: Decision Tree, Multi-Layer Perceptron neural network and the Automated Machine Learning process. Finally, the best model of each approach was applied to a given use case to analyze its performance in a possible real application. The experimental results indicate that the mobility models achieved an average accuracy of 91% and 74% for UFJF and KTH, in that order. For the number of users prediction approach, the UFJF scenario obtained an adjustment coefficient of 0.96 and the KTH scenario of 0.94, which can be classified as an effective result given its complexity. Finally, when analyzing the use cases, the results demonstrated that the UFJF mobility model provides an average accuracy of 87.15%, and for the number of users model, the maximum accuracy comes close to 95.92% when predicting the number of access points needed to serve the users connected at that moment. Similarly, the KTH campus scenario obtained an average accuracy of 70.63% for mobility, and a maximum accuracy of 97.28% for the model of the number of users connected to the network.