Thesis of Yann Berthelot

Efficient Deep Reinforcement Learning for Industrial Process Control

Saint-Gobain Research Paris and the Scool team wish to collaborate to develop fundamental research activitiesconcerning the advanced control of industrial processes. The specific objective of the thesis is to study the capabilities of Model-Based Reinforcement Learning approaches to meet practical requirements for industrial process control, and to achieve better performance than currently used control, such as predictive control. In practice, the control agents developed during the thesis will seek to meet the following objectives: - Take advantage of existing historical data for the system to be controlled (actuators and observables) and the physical principles governing their interactions (conservation of mass, energy, causality). - Learn the control strategy by balancing exploration and exploitation objectives. - Achieve a robust control with respect to the approximation of the transition and return functions in order to keep the process running correctly during the exploration phases. - To provide robustness guarantees for the process control and to integrate limitations related to the industrial context (limited exploration areas). - To be stable with respect to the measurement noise of the sensors and to be able to adapt to changes occurring over time (data drifts and concept drifts). - Take advantage of the training of a control agent on a production line to learn more efficiently on another line (transfer learning). In particular, the impossibility of having a perfect model of the production line implies a necessity for the control agent to learn the production line. If most of these objectives are generic and specific to the use of a Model-Based Reinforcement Learning approach for the control, some are directly linked to specific problems in the industrial specific to the industrial context. Consequently, an important part of the work carried out could be the object of results openly disseminated in order to whole research community benefit from the generic results. On the other hand, a part of the work will be more specific to the particular industrial case study, in order to test these methods in an environment as close as possible to reality.

Jury

M. Philippe PREUX Professeur des universités Université de Lille Directeur de thèse, Mme Ann NOWé Full professor Vrije Universiteit Brussel Examinatrice, M. Antoine CANAGUIER-DURAND Docteur Saint-Gobain Research Paris Co-directeur de thèse, M. Bruno SCHERRER Chargé de recherche Inria Rapporteur, M. Sylvain LAMPRIER Professeur des universités Université d'Angers Rapporteur, M. Riad AKROUR Inria Invité, M. Adam WHITE University of Alberta Invité.

Thesis of the team SCOOL defended on 19/06/2026