In-context self-supervised learning for stochastic optimal control with transformer architecture

Details

Supervisors

Jungers, Raphaël M.

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

enThis thesis explores the potential of Transformer-based models—originally developed for language tasks—as tools for controlling physical systems in uncertain and dynamical environments. Building on prior work showing that Transformers can learn estimation tasks from data alone, we extend this idea to control. First, we adapt a supervised framework where the model maps state trajectories to control inputs, showing that Transformers can infer control strategies from context. To overcome limitations in accuracy, we then introduce a self-supervised approach where the model learns control laws by interacting with the system and minimizing a cost function over time. Our findings highlight the promise of Transformers as flexible, data-driven controllers for complex, safety-critical systems.