Reward Observability and the Limits of Offline Checkpoint Selection in RSSM World Models

Name: Reward Observability and the Limits of Offline Checkpoint Selection in RSSM World Models
Uploaded: 2026-10-10T17:52:51.974911+00:00
Duration: 245 s
Description: The Composite Reward Observability Fraction (CROF) is a single score combining reward alignment and structural metrics to select optimal world models for model-based A2C policies. This selection method trains a policy that significantly outperforms model-free baselines using substantially fewer real-world interactions.

The Composite Reward Observability Fraction (CROF) is a single score combining reward alignment and structural metrics to select optimal world models for model-based A2C policies. This selection method trains a policy that significantly outperforms model-free baselines using substantially fewer real-world interactions.

Open the full episode