Diffusion models in offline reinforcement learning

Summary

Offline reinforcement learning (RL) trains decision-making agents from fixed datasets, without interacting with the environment during training. This thesis investigates how diffusion models can be integrated into offline RL, leveraging their ability to capture complex, multimodal distributions and to generate action or trajectory sequences via iterative denoising. We study diffusion-based methods in a real-world airplane trajectory dataset, focusing on goal-reaching constraints and generalization across varying dataset complexity. Overall, the findings indicate that diffusion models can generate feasible trajectories while accommodating domain-specific constraints, supporting their role as a flexible and robust approach for offline RL and constrained planning.