The student Haritz Odriozola Olalde an EXCELLENT with mention DOCTORATE INTERNATIONAL and INDUSTRIAL DOCTORATE
The student Haritz Odriozola Olalde an EXCELLENT with mention DOCTORATE INTERNATIONAL and INDUSTRIAL DOCTORATE
The student Haritz Odriozola Olalde an EXCELLENT with mention DOCTORATE INTERNATIONAL and INDUSTRIAL DOCTORATE
- Thesis title: Safe Reinforcement Learning-based adaptive control software platform for domain-shift environments
Court:
- Presidency: Matthias Althoff (Technische Universität München)
- Vocal: Viviane Cadenat (LAAS-CNRS)
- Vocal: Juan Ignacio Vazquez Gómez (Universidad de Deusto)
- Vocal: Aizea Lojo Novo (Ikerlan)
- Secretary: Tomaso Poggi (Mondragon Unibertsitatea)
Abstract:
Machine Learning (ML) is increasingly used in industrial applications, but its deployment in safety-critical contexts remains challenging due to technological and regulatory gaps. Reinforcement Learning (RL), known for its adaptability, is popular in sectors like robotics and autonomous vehicles, yet it shares similar safety concerns. Shielded RL offers safety guarantees by filtering unsafe actions using an abstract model, but domain shift problems can compromise these guarantees. This thesis introduces the Fear Field framework, inspired by intelligent life forms, to detect model inaccuracies and adjust agent conservativeness. Experimental results in simulated domain shift tasks show significant safety improvements, up to two orders of magnitude, while maintaining or enhancing performance and scalability. The work demonstrates that biologically inspired mechanisms can advance Shielded RL integration into safety-critical industrial systems.