Reproduce “Stein Variational Evolution Strategies” 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
Cornelius V. Braun, Robert T. Lange, Marc Toussaint
Stein Variational CMA-ES (SV-CMA-ES) is a gradient-free method that integrates Stein Variational Gradient Descent (SVGD) with Covariance Matrix Adaptation Evolution Strategy (CMA-ES) for diverse sampling and blackbox optimization. This approach replaces SVGD's gradient dependency with adaptive CMA-ES updates, leading to improved performance over other gradient-free SVGD variants on synthetic densities, Bayesian logistic regression, and reinforcement learning tasks while enhancing exploration and solution diversity.
Key insights
- CMA-ES's adaptive step-size and covariance matrix adaptation offer an effective, gradient-free proxy for the score function required by SVGD.
- Combining SVGD's particle repulsion with CMA-ES's adaptive exploration enhances the ability to navigate multimodal landscapes and generate diverse, high-quality solutions.
- A hybrid particle update, using individual CMA-ES mean updates for driving force and global SVGD repulsion, effectively balances local exploitation with global exploration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.