Hadrien HOURS - PhD Student
Date: - Location: Eurecom
Telecommunication networks are complex systems for which many factors act together to provide clients a service and a given quality. As Internet becomes more and more part of our daily life, its performance, along with its complexity and diversity, is in a constant evolution. The technologies supporting the evolution of the Internet are becoming more efficient, complex and expensive. Understanding the role of each parameter in the performance of such a system has become a tedious (if not impossible) task. Many operators still rely on back-of-the-envelope computations or pure domain knowledge to predict the effect of changing some components in their network, such solutions are neither reliable nor scalable. With the cost and risk that interventions on telecommunication networks represent, it is necessary to adopt a more scalable, automatic and formal approach, which is the topic of our work. In this talk, we will present an innovative approach to study telecommunication network performance based on causality. Causality has seen an important evolution in the recent years, particularly thanks to Judea Pearl (ACM Turing Award 2011, National Academy of science 2014) who participated in the development and formalism of a causal language along with the modeling of causal models through Bayesian networks. Causal models in combination with a Bayesian network representation offer a concise and simple visualization of complex systems. They also provide powerful theorems to predict the effect of interventions on these systems. While the theory is well developed its application is often limited by the constraints and complexity of the system one wants to study. Through an example of the study of FTP traffic, we will see how to adapt the different methods to our problem and develop a systematic approach that takes full advantage of the causal theory while taking into account the domain and resource constraints
