Inferindo a qualidade de serviço em redes via aprendizado por reforço

Abstract

The increase in the number of Internet users and the growth in the amount of traffic consumed by each of them, has generated in recent years, a demand for improvements in the quality of service (QoS) provided by Internet service providers (ISPs). The main objective of this work is to model the network conditions of an ISP and to infer if a change in these conditions will remain in the future. In this way, the ISP’s management system can be activated when it is predicted that the degradation of the network condition will remain for a certain time. A hidden Markov model (HMM) is defined from time series of packet loss rates collected on ISP’s home routers, and the model states are mapped to ISP network conditions. Then, a Reinforcement Learning technique is used in the time series of network conditions, to infer the performance of the network in the future and, possibly, indicate the triggering of a maintenance alarm. From the result of the Reinforcement Learning algorithm, the network manager can trigger maintenance actions or changes in network operating parameters. Results obtained using a real ISP dataset show good performance of the method through measures of accuracy, precision, recall and f1-score.