Reproduce “Reinforcement Learning with Structured Hierarchical Grammar Representations of Actions” 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
Petros Christodoulou, Robert Tjarko Lange, Ali Shafti, A. Aldo Faisal
AG-RL addresses reinforcement learning’s interaction bottleneck by inferring recurring action sequences from an agent’s behavior and adding them as selectable macro-actions. The authors test it with two off-policy Atari agents. After 100,000 environment steps, AG-SAC improves scores in 19 of 20 games versus SAC, with a median per-game improvement of 96%; this measures fixed-step scores, not wall-clock time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.