Use SVD instead of Eigendecomposition for greater numerical stability

srom / cma-es

Covariance Matrix Adaptation Evolution Strategy (CMA-ES)

MIT License

56 stars 13 forks source link

Use SVD instead of Eigendecomposition for greater numerical stability #2

Closed srom closed 4 years ago

srom commented 4 years ago

https://github.com/srom/cma-es/issues/1

Sometimes in my experiments the eigendecomposition fails when the initial state is chosen poorly. Wouldn't it make sense to use the SVD instead of the eigendecomposition?