Retraining strategy for diffusion models across generations

Determine whether to retrain a single diffusion model across successive generations or to train separate diffusion models from scratch at each generation using updated heuristic datasets within the HADES/CHARLES-D evolutionary optimization framework.

Background

In the Discussion, the authors outline limitations and practical considerations for their diffusion-model-based evolutionary strategies (HADES and CHARLES-D). A key unresolved design choice is how to manage the diffusion model across generations: either continually retrain one model with new heuristic data or repeatedly train fresh models from scratch using updated buffers.

This question directly impacts the algorithm’s memory, adaptability, and potential inductive biases introduced by the generative model, and the authors explicitly mark it as unresolved within their current framework.