Reproduce “How Training Data Impacts Performance in Learning-based Control” 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
Armin Lederer, Alexandre Capone, Jonas Umlauft, Sandra Hirche
Researchers at the Technical University of Munich developed an analytical framework to quantify how training data density impacts learning-based control performance using Gaussian Processes. The work introduces a "ρ-gap" measure to assess data quality and derives an ultimate tracking error bound that explicitly accounts for model uncertainty, demonstrating its utility in feedback linearization for nonlinear systems.
Key insights
- Gaussian Process posterior standard deviations provide a direct analytical link between training data density and the uncertainty components impacting control performance within a Lyapunov stability framework.
- The M-fill distance demonstrates that the effectiveness of training data for identifying different system functions (e.g., f(x) vs. g(x)) is dependent on the magnitude of associated control inputs.
- The ρ-gap effectively quantifies state-dependent deficiencies in training data density relative to the minimum requirements for achieving specified control performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.