Reproduce “Distributionally Robust Multi-Task Reinforcement Learning via Adaptive Task Sampling” 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
Nicholas E. Corrado, Wenyuan Huang, Josiah P. Hanna
DRATS reallocates interaction data across multi-task reinforcement learning tasks toward those furthest below desired returns, while retaining some sampling of every task. It estimates target returns online and adjusts task probabilities. Across MuJoCo6, MT10, and MT50, it reports higher aggregate and worst-task returns, but early underestimation can under-allocate data to tasks that still have room to improve.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.