How to optimize the hyperparameters?

[kitpeng11]

The printed out likelihood is negative after 50 optimization. Is there any way to get a set of good hyperparameters? Thanks

[mkrompiec]

You can try restarting from a different initial guess of hyperparameters. But the most likely reasons are that you don’t have enough training data, or the system can’t be adequately modelled by the kernel you chose.

[jonpvandermause]

Hi kitpeng11,

You can interpret the log marginal likelihood as a measure of how well your training data is described by the current model. As mkrompiec notes, a negative likelihood after hyperparameter optimization usually indicates that a more sophisticated model is needed.

Have you tried repeating the optimization with different kernels?

Best, Jon

[YuuuXie]

Close stale issue