Advances in Graph-Based Reinforcement Learning and Autonomous Decision-Making

The field of reinforcement learning and autonomous decision-making is moving towards the development of more robust and generalizable models. Researchers are exploring the use of graph-based methods to improve the reasoning capabilities of agents in complex environments. This includes the application of graph neural networks, transformers, and other techniques to capture local and global dependencies in graph-structured data. Another trend is the integration of uncertainty-aware decision-making frameworks, which can amplify learning from uncertain states and maintain stability across common transitions. Additionally, there is a growing interest in automating the discovery of useful abstractions directly from visual data, enabling more scalable and applicable planning frameworks in real-world robotic domains. Notable papers include:

  • Vejde, a framework for inductive deep reinforcement learning based on factor graph color refinement, which demonstrates impressive generalization capabilities.
  • GRATE, a graph transformer-based approach for time-efficient autonomous robot exploration, which exhibits better exploration efficiency than state-of-the-art baselines.
  • An Uncertainty-Weighted Decision Transformer for navigation in dense, complex driving scenarios, which consistently outperforms other baselines in terms of reward and behavioral stability.

Sources

Vejde: A Framework for Inductive Deep Reinforcement Learning Based on Factor Graph Color Refinement

GRATE: a Graph transformer-based deep Reinforcement learning Approach for Time-efficient autonomous robot Exploration

An Uncertainty-Weighted Decision Transformer for Navigation in Dense, Complex Driving Scenarios

Learning Discrete Abstractions for Visual Rearrangement Tasks Using Vision-Guided Graph Coloring

Leveraging Reinforcement Learning, Genetic Algorithms and Transformers for background determination in particle physics

Built with on top of