Description du projet
Rendre l’expérience numérique plus humaine
Il est primordial que les progrès technologiques soutiennent les activités et les interactions humaines dans les domaines des soins de santé, de la mobilité et des systèmes d’infrastructures. Par exemple, la mise en place de soins de santé plus humains exige la création d’interfaces numériques qui permettent des interactions plus humaines avec le système. C’est précisément l’objectif des systèmes anthropocentriques, dans lesquels l’homme est à la fois un élément du système de contrôle et un critère de conception. Le projet CO-MAN, financé par l’UE, va mettre au point un cadre pour un contrôle orienté données et adapté à l’utilisateur qui prévoit des garanties de performance. Le plus grand défi consistera à fusionner les techniques de modélisation probabiliste non paramétrique héritées de la théorie de l’apprentissage statistique avec des méthodes innovantes de contrôle sensibles au risque tout en incluant la modélisation active de l’utilisateur. L’aspect révolutionnaire tient aux efforts déployés pour obtenir un apprentissage mécanique fiable qui offre des résultats inédits en ce qui concerne les limites théoriques du comportement d’apprentissage.
Objectif
Many control systems of the future involve a tight interaction or even symbiosis with the human user. High-impact application domains of human-centric systems include healthcare, mobility, and infrastructure systems. In human-centric systems the human is both, an element of the control system, and a design criterion with individual requirements that need to be satisfied. Safety - despite the high uncertainty of human behavior - and maximization of individual user experience are the primary objectives for control design in human-centric systems. The visionary goal of CO-MAN is to contribute to the fundamental understanding and principled approach to the control of smart human-centric systems. We will develop a novel framework for user-adaptive data-driven control with performance guarantees in order to address the scientific challenges of high uncertainty and individual user requirements. The grand challenge is to unify the two previously separate paradigms of model-based control with its rigorous guarantees but limited modeling base and machine learning algorithms with its flexible modeling concepts but lack of guarantees. The breakthrough enabling idea is to merge probabilistic non-parametric modeling techniques from statistical learning theory with novel risk-aware control methodologies while including active user modeling. The game changer is the current push towards reliable machine learning with novel results on theoretical bounds for learning behavior. Because of favorable properties we will focus on Gaussian Processes to model user behavior and preferences and translate the naturally quantified model uncertainty into closed loop behavior guarantees through a confidence-driven human-interactive control approach. The PI is in a perfect position to achieve the envisioned goal of super-individualized data-driven control with performance guarantees given the highly visible preliminary results and leadership in the area of human-cyber-physical systems.
Champ scientifique
Not validated
Not validated
- engineering and technologyelectrical engineering, electronic engineering, information engineeringelectronic engineeringcontrol systems
- natural sciencesbiological sciencesbiological behavioural sciencesethologybiological interactions
- natural sciencescomputer and information sciencesartificial intelligencemachine learning
Programme(s)
Régime de financement
ERC-COG - Consolidator GrantInstitution d’accueil
80333 Muenchen
Allemagne