Reproduce “TaSIL: Taylor Series Imitation 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
Daniel Pfrommer, Thomas T.C.K. Zhang, Stephen Tu, Nikolai Matni
Researchers from the University of Pennsylvania and MIT introduce Taylor Series Imitation Learning (TaSIL), an off-policy method that reduces the imitation gap by matching higher-order derivatives of an expert policy. TaSIL improves sample efficiency across various MuJoCo tasks and provides theoretical guarantees linking expert robustness to the necessary order of derivative matching.
Key insights
- The required order `p` of Taylor expansion in TaSIL is directly determined by the decay rate of the expert's δ-ISS Class K function `γ(x)`; more robust experts require lower-order derivative matching.
- Matching higher-order derivatives effectively transfers the local dynamics and robustness of the expert, providing a mechanism to mitigate distribution shift without interactive data collection.
- Offline imitation learning can be PAC-learnable if the expert policy exhibits sufficient robustness, with TaSIL providing concrete conditions and an algorithm to achieve this.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.