Date depot: 6 avril 2021 Titre: Fast Deep Reinforcement Learning Techniques for self-configuration and optimization in IoT Directeur de thèse: Naceur MALOUCH (LIP6) Domaine scientifique: Sciences et technologies de l'information et de la communication Thématique CNRS : Non defini Resumé: Since IoT networks include a large number of connected devices that must be controlled, classic optimization and control algorithms are not capable anymore to control and to make efficient decisions. Indeed, not only the network is large-scaled but also the deployed protocols and softwares are becoming more and more complex. Classic algorithms reduce this complexity usually by fixing some parameters intuitively and/or by using strong assumptions regarding especially the objective functions. Deep Reinforcement Learning is a powerful tool that can be used in order to analyze network data and reach dynamically optimal configurations. Several previous works have shown the potential of DRL in different types of networks. The objective of the thesis is to explore how DRL can be applied and improved to provide faster decisions in IoT networks where the performance in terms of transmission, reliability and energy consumption is very sensitive to network parameters. This is thanks to the continuous interaction with the environment to learn the best action to perform. DRL can revolutionize optimization and control of future IoT networks. The PhD student will be provided some IoT devices and softwares so he can experiment concretely new DRL techniques. Before that, simulations environment will be used.