Covariance Matrix Adaptation

Covariance Matrix Adaptation (CMA) is an optimization technique that efficiently explores a search space by adapting a multivariate Gaussian distribution based on past search results. Current research focuses on extending CMA's capabilities to high-dimensional problems, particularly within Bayesian optimization and multi-task learning frameworks, often incorporating it into modular algorithms or using approximations to improve scalability. This work is significant because CMA's ability to effectively navigate complex search landscapes has broad applications, including improving the training of neural networks, optimizing robotic controllers, and enhancing the performance of black-box optimization methods.

Papers