Reproduce “Covariance Matrix Adaptation MAP-Annealing” 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
Matthew C. Fontaine, Stefanos Nikolaidis
Covariance Matrix Adaptation MAP-Annealing (CMA-MAE) introduces an annealed archive learning rate to enhance Quality Diversity (QD) optimization, improving exploration on flat objective landscapes and maintaining consistent performance across varying archive resolutions. The method empirically outperforms existing state-of-the-art derivative-free QD algorithms and its differentiable counterpart, CMA-MAEGA, demonstrates strong gains in Differentiable Quality Diversity tasks.
Key insights
- The annealing mechanism, through a learning rate (alpha) strictly between 0 and 1, offers a principled and robust trade-off between exploitation and exploration in QD algorithms.
- CMA-MAE performs "density descent" on flat objective functions, enabling continued exploration and archive filling even when objective values offer no direct guidance for improvement.
- The algorithm's performance becomes effectively invariant to archive resolution when the derived alpha conversion formula is applied, significantly reducing hyperparameter tuning complexity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.