Machine learning for predicting losses in the networks

In modern computer and telecommunication networks, traffics are transmitted based on packet-mode protocols. Due to the dynamic as well as the complicated properties of networking and the spread of broadband and internet access, one of the major issues is packet loss. Some of the protocols, like Transmission Control Protocol (TCP), can tackle this problem and recover losses, but for Real-time Transport Protocol (RTP), packet delivery cannot be guaranteed, which will significantly affect the Quality of Experience (QoE) for services. Therefore, on the one hand, a preventive solution that can forecast the packet loss is needed to help build an adaptive mechanism and improve the QoE, and on the other hand, studies of packet loss can help to understand the network status, e.g., congestion. In this thesis, we will focus on RTP-based Real-time Communication (RTC) services, and comprehensive analyses for a case study based on two RTC applications will be conducted to understand the specific properties of packet loss. On top of that, several machine learning approaches will be developed to predict losses based on historical statistics of packets. In particular, an improvement to the command-line tool Retina will be made to derive the exact number of packet losses, and then, the dataset about statistics of RTP packets will be analyzed, processed, and utilized to feed the machine learning models. Finally, in-depth studies will be performed as well.