Decentralized learning and its industrial applications

In many industrial control problems, several agents interact in a shared environment and may be subject to certain constraints. These constraints are often imposed to guarantee the safety of the physical system. For example, in a wind farm, the turbines must satisfy certain constraints to avoid too much load which can harm their physical structure. Despite its importance, taking constraints into account in the multi-agent reinforcement learning framework (MARL for Multi-Agent RL, RL for Reinforcement Learning) has only emerged in recent years. The objective of this thesis is to develop new MARL algorithms that can satisfy security constraints, in the case where agents collaborate and execute actions in a decentralized manner. On the one hand, for a better scalability, we want to make the algorithms as decentralized as possible. A study of the convergence guarantees of the algorithms is envisaged to achieve a theoretical understanding of the performance and/or limitations of the solutions developed. On the other hand, we aim to arrive at simple algorithms that can be applied to other similar problems, with little or no parameter tuning. To validate theoretical developments, we wish to address the challenges posed by the control of wind farms, in particular, the control of a wind farm to maximize its total production while satisfying safety constraints on the turbines.

Keywords: decentralized learning, reinforcement learning, multi-agent

• Academic supervisor    Prof. Ana BUSIC,  INRIA Paris / DI ENS, Université PSL, ORCID : 0000-0002-4133-3739
• Doctoral School    ED386, DI ENS
• IFPEN supervisor    Dr Jiamin ZHU, ORCID : 0000-0002-4552-5519
• PhD location    IFPEN, Rueil-Malmaison & INRIA, Paris
• Duration and start date    3 years, starting in the fourth quarter 2024 
• Employer    IFPEN
• Academic requirements    University Master degree in mathematics or computer and information sciences
• Language requirements    English level B2 (CEFR)
• Other requirements    Programming skills: Python, Matlab/Simulink

To apply, please send your cover letter and CV to the IFPEN supervisor indicated below.