Experiments

[Sceki]

• different time lags & resolutions
• different inputs
• comparison between simpler (lstm) and more complex (TFT) NN architectures
• hyperparameter search (see Hyperparameter search for NN training #5 )
• comparison at specific hold-out sets vs TIE-GCM and possibly other physics-based models

[Sceki]

For a comprehensive list, see: https://docs.google.com/document/d/1JxhYZ273vo02mVrBW5IrYdKohYGrQtvNgc2oWG-bOxs/edit

[Sceki]

Most important part of the experiments done and is now on W&B. Currently, waiting to analyze those for the final presentation.

• the NN hyperparameter search is not priority for now
• the TIE-GCM comparison does not seem possible (the runs we sent are not being processed in their servers)

So turning this on hold for now.