Koopman-Operator-based Reinforcement Learning Control of Partial Differential Equations

Descripción del proyecto

Control robusto y de baja energía mediante el aprendizaje por refuerzo de sistemas dinámicos a gran escala

Mejorar el análisis de macrodatos, que consume mucha energía y requiere un gran esfuerzo informático, es un objetivo fundamental en muchos campos, desde la investigación básica sobre los orígenes del universo hasta la predicción del cambio climático o el control de sistemas dinámicos a gran escala en aplicaciones de ingeniería fundamentales para la seguridad. El aprendizaje por refuerzo (AR) es una técnica prometedora para reducir los requisitos energéticos de la horma. Sin embargo, los métodos lineales o de núcleo no se adaptan bien, lo cual minimiza su utilidad para las aplicaciones prácticas. Además, requieren muchos cálculos y, por lo tanto, mucha energía. El proyecto KoOpeRaDE, financiado por el Consejo Europeo de Investigación, pretende aprovechar la teoría del control, la teoría de la aproximación y el aprendizaje automático para reducir la complejidad y apoyar el desarrollo de controladores del AR innovadores para sistemas de ingeniería a gran escala con garantías de rendimiento.

Objetivo

An unprecedented energy crisis is looming over us. In order to transition to a greener and more energy-efficient society, existing technologies need to be improved and novel techniques such as nuclear fusion developed. This requires the stabilization of aerodynamics, heat transfer or combustion and fusion processes and thus, the development of efficient control strategies for large-scale dynamical systems. In recent years, reinforcement learning (RL) has emerged as a highly promising data-driven technique. Unfortunately, we cannot trust RL to handle our most important and complex systems, since the resulting controllers do not possess performance guarantees. Certifiable RL approaches such as linear or kernel methods tend to scale poorly, such that their applicability is limited to toy examples. In contrast to other application areas, this is a complete show-stopper for safety-critical engineering. Moreover, the training is extremely data hungry and costly, due to which RL itself contributes to the energy crisis.
The vision of this project is to develop new foundational methods to equip RL controllers for large-scale engineering systems with performance guarantees by exploiting system knowledge and systematically reducing the complexity. To achieve this, I will target three major breakthroughs, consisting of (A) global linearization of the dynamics via the Koopman operator framework, (B) the extension of certified Q-learning to continuous action spaces via control quantization, and (C) the detection and exploitation of symmetries in the system dynamics.
The project requires significant joint advancements in several challenging areas such as control, approximation theory and machine learning. In the case of success, the resulting controllers will provide a massive advancement of RL towards safety-critical engineering applications and significantly contribute to the challenge of meeting the future energy demands of our society.