search

mlr-org / mlrMBO

Toolbox for Bayesian Optimization and Model-Based Optimization in R
https://mlrmbo.mlr-org.com
Other
187 stars 47 forks source link

When to save state on disc? #105

Open danielhorn opened 9 years ago

danielhorn commented 9 years ago

While reading issue #102 this came to my mind:

At the moment, we save the optimization starte after the evaluation of the target function. Wouldn't it be more useful to save the state befor the function evaluation (after the proposal of new points)?

The reason is: From my experience with my experiments - most crashes appear during the evaluation of the target function. And, a restart from the save file starts with proposing new points. But - the point proposal was allready done before the crash. It just was not saved. The proposal of new points can take several minutes, why should it be done again? Wouldn't it be better so save the state after the points were proposed?

Okay, I see: If propPoints crashes, the last function evaluations have to be done again, and this is even more expensive.

Perhaps a soluttion would be to save both after propPoints and evalTargetFunction?

berndbischl commented 9 years ago

Yes, we should probably save after both operations, proposal and eval. But stuff like this is always complicated. I will check.