Covariance Matrix Adaptation-Evolution Strategy (CMA-ES)
Used in the Game-Theoretic Objective Space Planning, Hongrui Zheng introduced me to this idea.
It’s used in the World Models paper.
GA vs. Vanilla ES vs. CMA-ES?
- GA: “Mix and match DNA” — relies on recombination + random bit flips.
- Vanilla ES: “Spray random Gaussian noise in every direction equally” — works but struggles if the landscape is skewed.
- CMA-ES: “Learn the shape of the landscape” — reshapes its search cloud into an ellipse that aligns with valleys and ridges, zooming in efficiently.