Description
Date depot: 7 avril 2023
Titre: Model-based Reinforcement Learning with a known physical prior
Directeur de thèse:
Olivier SIGAUD (ISIR (EDITE))
Directeur de thèse:
Nicolas THOME (ISIR (EDITE))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: Model-based Reinforcement Learning (MBRL) is a control optimization technique used when the model of the system to be controlled is either known or learnt online from interaction data. Recent techniques based on Neural Ordinary Differential equations (Neural ODEs) have shown the advantages of combining partial knowledge of the dynamics of a system with data-driven learning techniques that learn the residual between the known model and what the data tell us about it. The objective of this PhD thesis is to leverage Neural ODEs in the context of MBRL to learn a model that is more accurate than what the prior says, and that is learnt faster than learning without the prior.