Reproduce “Discovering Representations for Black-box 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
Adam Gaier, Alexander Asteroth, Jean-Baptiste Mouret
A new algorithm, DDE-Elites, automates the discovery of effective solution representations for black-box optimization by combining Quality Diversity methods with Variational Autoencoders. This approach efficiently scales optimization to high-dimensional problems, such as a 1000-joint robot arm inverse kinematics task, and the learned Data-Driven Encodings (DDEs) serve as reusable representations that accelerate subsequent optimization and inherently embed desirable solution characteristics.
Key insights
- Quality Diversity algorithms are well-suited for generating diverse, high-performing datasets from which generative models can learn effective and biased representations.
- Continuously updating the learned representation (DDE) from the evolving optimization archive creates a virtuous cycle that enhances both the representation and the optimization process.
- A learned low-dimensional encoding can make high-dimensional black-box problems more tractable for optimization algorithms by providing an inherently biased and expressive search space.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.