Diffusion-Augmented Markov Decision Processes for Maximum Entropy Reinforcement Learning

Name: Diffusion-Augmented Markov Decision Processes for Maximum Entropy Reinforcement Learning
Uploaded: 2026-10-02T09:34:35.080908+00:00
Duration: 689 s
Description: This work introduces Diffusion-Augmented Markov Decision Processes (DA-MDPs) to extend Maximum Entropy Reinforcement Learning. It uses diffusion models to sample optimal action trajectories by minimizing a tractable upper bound on the reverse KL divergence. This allows for modeling complex, multimodal action distributions that standard methods cannot capture.

This work introduces Diffusion-Augmented Markov Decision Processes (DA-MDPs) to extend Maximum Entropy Reinforcement Learning. It uses diffusion models to sample optimal action trajectories by minimizing a tractable upper bound on the reverse KL divergence. This allows for modeling complex, multimodal action distributions that standard methods cannot capture.

Open the full episode