Reproduce “Training Stability under Adaptive Curricula” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Johan Obando-Ceron, Walter Mayor, Michael Dennis, Aaron Courville, Pablo Samuel Castro
The authors test whether stabilizing reinforcement-learning learners improves adaptive curricula without changing their rules. In PPO agents trained on Kinetix, weight clipping and a categorical value-prediction method improve learning across tested curriculum mechanisms. Under prioritized level replay, stabilization also helps with longer training and parallelism settings. Evidence is limited to PPO and Kinetix; the interventions do not establish a single cause.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.