The current demo uses maximum likelihood estimation for hyperparameters, but it fails sometimes. Especially, the length-scale hyperparameter in the ARD squared exponential kernel becomes unrealistically small, resulting in a "no relevance" kernel matrix.
A possible fix is to use maximum a posteriori estimation with an appropriate prior assumption for the hyperparameters.
The current demo uses maximum likelihood estimation for hyperparameters, but it fails sometimes. Especially, the length-scale hyperparameter in the ARD squared exponential kernel becomes unrealistically small, resulting in a "no relevance" kernel matrix.
A possible fix is to use maximum a posteriori estimation with an appropriate prior assumption for the hyperparameters.