search

microprediction / humpday

Elo ratings for global black box derivative-free optimizers
MIT License
132 stars 18 forks source link

scipy errors #3

Closed rkern closed 3 years ago

rkern commented 3 years ago

https://github.com/microprediction/humpday/blob/21969d33bffdbafc7a399f64e29e8a5e483fc28b/humpday/optimizers/scipycube.py#L28

Currently, you are always running method='powell' regardless of which method you've requested. The options dict is for arguments specific to the optimization method, typically tolerances. There is no maxfev option, only maxiter.

The dogleg method requires a Jacobian function to be provided; it is not a derivative-free method.

The SLSQP method is misspelled as SLQSP, here and in shgocube.

It will be interesting to see how fixing these affects the ELO ratings.

microprediction commented 3 years ago

Fantastic catches. Thanks so much.

microprediction commented 3 years ago

And yes, we'll see how the new Elo ratings are looking by hump day.

microprediction commented 3 years ago

Pushed those fixes, I think. Thanks again, Robert.