globus-labs / cascade

Fast and reliable dynamic simulations by concurrently learning, auditing, and using surrogates
MIT License
1 stars 0 forks source link

Better handle uncertainty with pretrained models #51

Open WardLT opened 3 months ago

WardLT commented 3 months ago

The current implementation of Proxima will not use an ML model if the ensemble of models produce 0 uncertainty for everything, which is the case when we start with a single pre-trained models (zero deviation between the predictions of the same model). That means that the model will never be used even if it is excellent.

Maybe we should assign a small positive value for the UQ even if there is a value of 0 so that it is possible for the model to run if the controller decides.