CMA-ES and Modern Continuous Evolutionary Search
Covariance Matrix Adaptation
Why CMA-ES Matters
CMA-ES is a powerful evolution strategy that adapts the covariance structure of the search distribution. This lets it learn correlations between variables and move efficiently through difficult continuous landscapes.
Search Distribution
Instead of mutating candidates independently, CMA-ES samples from a multivariate normal distribution , where is the mean, is the global step size, and is the covariance matrix.
Key Parameters
High-Level CMA-ES Pseudocode
-
1
Step 1: Initialize mean, covariance, and step size.
-
2
Step 2: Sample offspring from a multivariate normal distribution.
-
3
Step 3: Evaluate offspring fitness.
-
4
Step 4: Select the best offspring.
-
5
Step 5: Update the mean toward selected solutions.
-
6
Step 6: Update covariance to reflect successful search directions.
-
7
Step 7: Adapt step size and repeat.
CMA-ES vs Basic ES
Basic ES
- Uses simple Gaussian mutations
- May use fixed or self-adaptive step sizes
- Does not explicitly model correlations
CMA-ES
- Learns covariance structure
- Handles ill-conditioned problems well
- Often state-of-the-art for black-box continuous optimization
What additional structure does CMA-ES learn beyond a single step size?
The covariance matrix captures variable correlations and search shape.
Correct answer: A covariance matrix
Why can covariance adaptation be useful?
This helps the algorithm move along promising contours.
Correct answer: It captures correlations between variables and improves search direction.
CMA-ES is most naturally used for which type of optimization?
CMA-ES is designed for continuous, derivative-free optimization.
Correct answer: Continuous black-box optimization
Practical Note
CMA-ES is especially effective when variables interact strongly, the landscape is noisy, or scaling differs across dimensions.
Computation Cost
Maintaining and updating a covariance matrix can be expensive in very high dimensions.