Physics-informed reinforcement learning for complex environments, using graph neural networks

Overview

Neglecting to incorporate physics information into world models for reinforcement learning leads to reduced adaptability to dynamic and complex environments and overall learning outcomes.

In this project, we endeavour to develop and implement learnable models in reinforcement learning (RL) based on graph neural networks (GNNs). These models will integrate object and relation-centric representations to enable accurate predictions, strong generalization, and system identification in complex, dynamical systems. Additionally, we will focus on leveraging extensive world knowledge or physics information to refine representations of world models, thus enhancing learning outcomes and predictive capabilities.

Research activities

• Familiarise with RL setting and setup training framework.
• Design and learn world model from data and incorporated physics information (e.g. using GNNs).
• Evaluate performance, write, publish, and present research outcomes.

Outcomes

The aim of the project is to design improved learnable world models for reinforcement learning by incorporating physics information using graph neural networks (GNNs).

Skills and experience

You must have:

• a strong math background
• programming experience in Python
• some machine learning and/ or reinforcement learning experience is desired.

Keywords

• physics informed machine learning
• reinforcement learning
• graph neural networks
• GNN

Contact

Contact the supervisor for more information.