Master's Thesis - Enhance Performance of Neural-Network-Based Action Masking

17.08.2025, Studentische Hilfskräfte, Praktikantenstellen, Studienarbeiten

Provably safe reinforcement learning is critical for real-world safety-critical applications. One of the core challenges is to ensure that the agent does not take unsafe actions during both training and deployment. Action masking is a common technique to prevent the agent from selecting unsafe actions. Current methods often rely on hand-crafted rules or heuristics to define and compute safe actions, which can be conservative and difficult to scale. Neural networks have shown promise in learning to mask unsafe actions directly from data and then be used for training safe reinforcement learning agents. However, the performance of neural-network-based action masking is limited especially in complex and dynamic environments.

In this thesis, we aim to enhance the performance of neural-network-based action masking for reinforcement learning. The goal is to improve and extend the existing pipeline for neural-network-based action masking, implement and test curriculum learning techniques, and finally evaluate the performance of the enhanced action masking network in an autonomous driving scenario based on CommonRoad and CommonRoad-RL.

This thesis offers an opportunity to engage in practical applications of autonomous driving. The project also aims for a publication in a peer-reviewed conference or journal.

Your tasks:
- Familiarize with our current action masking techniques.
- Familiarize with the existing code base for neural-network-based action masking in CommonRoad-RL.
- Enhance the efficiency and performance of the existing action masking pipeline.
- Implement curriculum learning techniques to improve the performance of the action masking method.
- Evaluate the performance in an autonomous driving scenario.
- Documentation of your results.

Required skills:
- Knowledge of Reinforcement Learning and Curriculum Learning.
- Good Python programming skills and experience with PyTorch.

Please find the attached PDF for a detailed topic description.

If you are interested in this topic, please send an email to shuaiyi.li@tum.de with your CV and transcript with title "[Bachelor/Master Thesis Application] ..." :D

Kontakt: shuaiyi.li@tum.de