Reproduce “To Distill or Decide? Understanding the Algorithmic Trade-off in Partially Observable Reinforcement Learning” 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
Yuda Song, Dhruv Rohatgi, Aarti Singh, J. Andrew Bagnell
The authors study whether privileged expert distillation can train image-based controllers when observations hide state such as velocity. Their theory predicts that deterministic dynamics can let observations resolve uncertainty, while stochastic dynamics may favor reinforcement learning (RL) with longer histories. In simulated locomotion with injected motor noise, RL outperforms DAgger at higher noise; this is not evidence from real robots.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.