This study compares dynamic training strategies across encoders, critics and actors, and examines their interaction with network architecture. It argues for module-specific training choices rather than applying one sparse-training rule uniformly to every component. The analysis distinguishes topology changes from architectural changes as contributors to improved network scalability.

The linked preprint is also available under its earlier title, Rethinking the Role of Dynamic Sparse Training for Scalable Deep Reinforcement Learning.

Publication record.