Skip to main content
Vai all'homepage della Commissione europea (si apre in una nuova finestra)
italiano it
CORDIS - Risultati della ricerca dell’UE
CORDIS

Graphical models strike back: non-linearity, actions, and calibration

Obiettivo

The recent success of machine learning (ML) has been fueled by black-box supervised learning on massive amounts of data. While this has led to many applications throughout science, engineering, and industry, new areas such as healthcare, transport, or robotics require interactions between predictions and decisions from high-dimensional data, such as in reinforcement learning (RL), and provably well-calibrated uncertainty estimates.

This project aims to address these requirements with the same reliability and efficiency as supervised learning, which current RL algorithms do not possess. This will be achieved by opening the black box at an appropriate granularity and proposing a flexible, unified algorithmic framework based on probabilistic graphical models. The revival of this classical modeling tool presents significant, radically novel scientific challenges when dealing with high-dimensionality and nonlinearity: (1) ML must interact with partially observed environments or other agents; (2) theoretical guarantees are essential but remain largely underdeveloped, both a priori, to characterize the amount of resources needed for a particular problem, and a posteriori, to certify the performance of an algorithm on unseen data.

So-called influence diagrams have been precisely defined to extend graphical models by adding action and utility nodes. Their development, however, has so far been limited to the simplest white-box situations where all components of the learning systems have to be explicitly modeled. In order to tackle modern applications, we will leverage and extend the latest tools from ML and optimization, such as score-based generative modeling and sum-of-squares optimization. These new frameworks open a unique window of opportunity for a gray-box approach that models data at the right level.

This project will result in a suite of learning algorithms and open-source software coming with theoretical guarantees and a wide range of applications.

Campo scientifico (EuroSciVoc)

Parole chiave

Programma(i)

Argomento(i)

Meccanismo di finanziamento

HORIZON-ERC - HORIZON ERC Grants

Vedi tutti i progetti finanziati nell’ambito di questo schema di finanziamento

Invito a presentare proposte

(si apre in una nuova finestra) ERC-2025-ADG

Vedi tutti i progetti finanziati nell’ambito del bando

Istituzione ospitante

INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET AUTOMATIQUE
Contributo netto dell'UE
€ 2 498 920,00
Indirizzo
DOMAINE DE VOLUCEAU ROCQUENCOURT
78153 Le Chesnay Cedex
Francia
Regione
Ile-de-France Ile-de-France Yvelines
Tipo di attività
Research Organisations
Collegamenti
Costo totale
Nessun dato

Beneficiari (1)