Techniques d'Apprentissage par Renforcement pour le Routage Adaptatif dans les Réseaux de Télécommunication à Trafic Irrégulie

The aim of this thesis is to propose an algorithmic approach, which allows to treat the problems of adaptive routing (AR) in telecommunication networks with irregular traffic. The analysis of the existing approaches has lead us to base our work on the Q-Routing (QR) algorithm. This algorithm uses a reinforcement learning technique which is based on Markov models. The efficiency of these routing approaches depends on information about the network load and the nature of data flows. This information must be sufficient and relevant and it has to reflect the real network load during the decision making phase. To overcome drawbacks of techniques using QR, we have proposed two AR algorithms. The first one, which is called Q-Neural Routing, is based on a stochastic neural model, used for parameter estimation and updating required for routing. In order to reduce the convergence time, a second approach is proposed: k-Shortest path Q-Routing. It is based on a multi-paths routing technique combined with the QR algorithm. In this case, the exploration space is limited to k-Best paths. The proposed algorithms are validated and compared to traditional approaches using the OPNET Simulator. Their efficiency, with respect to AR, is illustrated. In fact, these algorithms allow taking into account the network state in a better way than the classical approaches do.