The key in init_params for combination kernels are currently reused within each internal kernel. As such, if one uses multiple of the same kernel within (e.g. to use a sum of kernels with multiple lengthscales), they will be initialised to be identical. This is problematic as the symmetry between them forces them to stay the same using gradient-based local optimization.
This splits the key so that each kernel will be initialised to its own state (if using random initialisation).
The key in init_params for combination kernels are currently reused within each internal kernel. As such, if one uses multiple of the same kernel within (e.g. to use a sum of kernels with multiple lengthscales), they will be initialised to be identical. This is problematic as the symmetry between them forces them to stay the same using gradient-based local optimization.
This splits the key so that each kernel will be initialised to its own state (if using random initialisation).