Improving in-context generalization in reinforcement learning through asymmetric model-based methods using adequate representation and exploration techniques

Project Information

GenRL

Grant agreement ID: 101284668

DOI

10.3030/101284668

EC signature date 23 February 2026

Start date 1 April 2026

End date 30 June 2028

Funded under

• Marie Skłodowska-Curie Actions (MSCA)

Total cost

No data

EU contribution

€ 238 388,40

Coordinated by

UNIVERSITE DE LIEGE
Belgium

Project description

Improving the generalisation capabilities of reinforcement learning

Reinforcement learning (RL) is a highly promising framework for solving decision-making problems, particularly because it requires few assumptions about the problems it is used for, which allows the algorithm to learn near-optimal behaviours even in unknown environments. However, the empirical successes of RL algorithms in the real world have often required extensive dedicated training and have shown limited adaptability, in contrast to the significant generalisation capabilities of modern large generative models. Supported by the Marie Skłodowska-Curie Actions programme, the GenRL project will leverage various methods to improve generalisation and efficiency. In particular, the project will take advantage of key progresses in representation learning, asymmetric learning and model-based RL, thereby advancing the reliability of RL in real-world applications.

Objective

Reinforcement learning (RL) is an appealing framework for solving decision-making problems, notably because it makes few assumptions about the problem at hand. In its purest form, the promise of an RL algorithm is to learn an optimal behavior from interaction with an unknown environment. There has been a plethora of empirical successes in real-world applications ranging from games to robotics. However, most of these achievements have required a dedicated training, and the learned behaviors have demonstrated limited generalization abilities. Compared to the generalization capabilities recently obtained in generative modeling with large pretrained models, notably in language generation, it appears clearly that RL has much room for improvement when it comes to generalization. We define generalization as the ability to maintain performance by adapting to environment changes (perception, dynamics, or rewards) based on the observable context only. Interestingly, in-context generalization is known to be equivalent to the problem of optimally controlling a partially observable environment. Coincidentally, the RL field has been bursting with discoveries over the last few years, with notable progresses in several domains that are closely related to generalization and partial observability: representation learning, model-based RL, asymmetric RL, and exploration. These advances inspire enthusiasm about the future of RL, and in particular about the idea of developing RL algorithms able to learn behaviors that generalize well. It motivates this research project that will improve generalization by (i) developing new world model architectures for model-based RL with effective imagination, (ii) developing asymmetric representation learning objectives for better convergence and sample-efficiency, (iii) designing suitable exploration strategies relying on the aforementioned representations, and (iv) benchmarking generalization on a real-world application, tertiary voltage control.

Fields of science (EuroSciVoc)

• natural sciences › computer and information sciences › artificial intelligence › machine learning › reinforcement learning
• engineering and technology › electrical engineering, electronic engineering, information engineering › electronic engineering › robotics

Keywords

• Reinforcement Learning
• In-Context Generalization
• Partially Observable Markov Decision Process
• Asymmetric Learning
• World Models
• Zero-Short Learning
• Representation Learning
• Model-Based
• RL
• POMDP

Programme(s)

• HORIZON.1.2 - Marie Skłodowska-Curie Actions (MSCA) MAIN PROGRAMME
  

Topic(s)

• HORIZON-MSCA-2025-PF-01-01 - MSCA Postdoctoral Fellowships 2025
  

Funding Scheme

HORIZON-TMA-MSCA-PF-GF - HORIZON TMA MSCA Postdoctoral Fellowships - Global Fellowships

Call for proposal

HORIZON-MSCA-2025-PF

Coordinator

UNIVERSITE DE LIEGE

• Net EU contribution: € 238 388,40
• Address: PLACE DU 20 AOUT 7 ; 4000 LIEGE ; Belgium
• Region: Région wallonne › Prov. Liège › Arr. Liège
• Activity type: Higher or Secondary Education Establishments
• Total cost: No data

Partners (1)

Partner

ROYAL INSTITUTION FOR THE ADVANCEMENT OF LEARNING MCGILL UNIVERSITY

Canada

• Net EU contribution: € 0,00
• Address: 845 SHERBROOKE STREET WEST ; H3A0G4 Montreal
• Activity type: Higher or Secondary Education Establishments
• Total cost: No data