Identify and evaluate alternative priority schemes for SCLD’s replay buffer

Investigate alternative prioritization schemes for the prioritized replay buffer used in training Sequential Controlled Langevin Diffusion (SCLD), including prioritization by importance weights via Radon–Nikodym derivatives, and determine their impact on training stability, sample efficiency, and sampling performance.

Background

The SCLD training procedure employs a prioritized replay buffer to stabilize optimization and improve sample efficiency; in the presented implementation, prioritization uses Radon–Nikodym derivative–based weights.

The authors explicitly note that many alternative prioritization strategies could be used and leave the exploration of these alternatives to future work, indicating an open design choice with potential impact on training dynamics and final performance.