Reproduce “Environment Shaping in Reinforcement Learning using State Abstraction” 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
Parameswaran Kamalaruban, Rati Devidze, Volkan Cevher, Adish Singla
The authors present a reinforcement-learning framework that compresses states with an expert-supplied abstraction and policy, then reshapes rewards, transitions, or both to favor that policy. A guarantee links near-optimality in the shaped environment to performance in the original task. In a gathering-game test, the representation shrinks from 15^4 states to 15^2, with curves improving over two baselines on this single designed task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.