Reproduce “Local Entropy Search over Descent Sequences for Bayesian Optimization” 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
David Stenger, Armin Lindicke, Alexander von Rohr, Sebastian Trimpe
Local Entropy Search (LES) is introduced as a Bayesian optimization framework that efficiently identifies local optima by maximizing information gain about the full descent sequence of an iterative optimizer applied to Gaussian Process samples. This method achieves superior sample efficiency compared to global and other local BO approaches in high-dimensional, complex black-box optimization problems.
Key insights
- Focusing Bayesian optimization on the *descent sequence* of an iterative optimizer, rather than a single global optimum, significantly improves sample efficiency for local search in high-dimensional problems.
- An information-theoretic approach that explicitly models and reduces uncertainty about the entire local optimization trajectory provides a more principled and efficient strategy than traditional local BO or global BO methods.
- The performance and behavior of LES are influenced by the choice of the internal local optimizer (e.g., ADAM, Gradient Descent, CMA-ES), suggesting adaptability to different problem characteristics and complexities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.