It would be nice to have the option to initialise multilevel sampling with some kind of emulator at the coarsest level, using a cold start. That is, only train an emulator after some samples have been collected on the higher levels, and then automatically turn on the emulator on the coarsest level once enough samples have been gathered. Then, according to some user-specified period, retrain the emulator according to the samples on the higher levels.
It is sort of tempting to add a bunch of emulator templates in e.g. pytorch, but I would rather not include such an extensive framework as a dependency. What I would instead suggest is create a wrapper that would accept any kind of pre-defined emulator and equip that wrapper with a simple interface consisting of .fit(), .update() and .evaluate() methods. That would allow the user to exploit their favourite emulator framework for their model and give a lot more flexibility.
It would be nice to have the option to initialise multilevel sampling with some kind of emulator at the coarsest level, using a cold start. That is, only train an emulator after some samples have been collected on the higher levels, and then automatically turn on the emulator on the coarsest level once enough samples have been gathered. Then, according to some user-specified period, retrain the emulator according to the samples on the higher levels.
It is sort of tempting to add a bunch of emulator templates in e.g.
pytorch, but I would rather not include such an extensive framework as a dependency. What I would instead suggest is create a wrapper that would accept any kind of pre-defined emulator and equip that wrapper with a simple interface consisting of .fit(), .update() and .evaluate() methods. That would allow the user to exploit their favourite emulator framework for their model and give a lot more flexibility.