Skip to main content
Aller à la page d’accueil de la Commission européenne (s’ouvre dans une nouvelle fenêtre)
français fr
CORDIS - Résultats de la recherche de l’UE
CORDIS

Ergodicity-aware continual reinforcement learning

Objectif

Current reinforcement learning (RL) algorithms optimize ensemble averages across episodes — a fundamental mismatch with real-world deployment, where we care about the single trajectory that actually unfolds. Real-world systems are continual: they evolve without reset, with irreversible consequences, and under non-stationary dynamics. This project develops the mathematical foundations and algorithms to move beyond expectation-based optimization toward learning systems that optimize what matters: individual trajectory performance in non-resetting, evolving environments.

Current RL fails on two fronts: it optimizes expected values that diverge from individual trajectory outcomes when ergodicity breaks, and it learns static policies that cannot adapt to environmental evolution. This project addresses both challenges simultaneously through ergodicity-aware continual RL, combining trajectory-centric objectives with adaptive, context-dependent policies. This enables autonomous vehicles adapting to seasonal changes over years, manufacturing systems evolving with production demands without downtime, and medical monitoring personalizing to patients over decades — applications infeasible with current RL theory.

Our contributions are threefold: (i) we clarify, when and why expected values fail, and derive mathematical principles for learning ergodicity transformations that restore meaningful optimization objectives in non-ergodic environments; (ii) we design algorithms that integrate these new criteria with adaptive policies, enabling agents to sustain performance across long horizons without resetting assumptions; (iii) we validate the framework in novel, non-episodic, and non-stationary benchmark environments. Together, these contributions provide the missing foundations for continual RL, replacing expectation-based reasoning with tools that optimize long-term growth and resilience of deployed agents.

Champ scientifique (EuroSciVoc)

Ce projet n'a pas encore été classé par EuroSciVoc.
Proposez les domaines scientifiques qui vous semblent les plus pertinents et aidez-nous à améliorer notre service de classification.

Mots‑clés

Programme(s)

Thème(s)

Régime de financement

HORIZON-ERC - HORIZON ERC Grants

Voir tous les projets financés dans le cadre de ce programme de financement

Appel à propositions

(s’ouvre dans une nouvelle fenêtre) ERC-2026-STG

Voir tous les projets financés au titre de cet appel

Institution d’accueil

AALTO KORKEAKOULUSAATIO SR
Contribution nette de l'UE
€ 1 498 318,00
Adresse
OTAKAARI 1
02150 Espoo
Finlande
Région
Manner-Suomi Helsinki-Uusimaa Helsinki-Uusimaa
Type d’activité
Higher or Secondary Education Establishments
Liens
Coût total
€ 1 498 318,00

Bénéficiaires (1)