Fitting is improved by proper data normalization and initialization. This is explained in the ex10/README.md and implemented together with some smaller fixes/changes into the FFNN lib and the example.
Contrary to the fitting example in NFM lib, this example concerns NN fitting via GSL's fit routines, with Levenberg-Marquardt solver. Compared to the NFM example, much better fits could be achieved here.
Adding the data normalization improved the stability. Next thing to check for is alternative cost functions / regularization.
Part 1: Trello card https://trello.com/c/MNffP2IG/16-try-to-understand-why-some-nn-fits-do-not-work-initial-nn-parameters
Fitting is improved by proper data normalization and initialization. This is explained in the ex10/README.md and implemented together with some smaller fixes/changes into the FFNN lib and the example.
Part 2: Trello card https://trello.com/c/QMRZ4r3u/14-use-gsl-lib-for-a-nn-fit
Contrary to the fitting example in NFM lib, this example concerns NN fitting via GSL's fit routines, with Levenberg-Marquardt solver. Compared to the NFM example, much better fits could be achieved here. Adding the data normalization improved the stability. Next thing to check for is alternative cost functions / regularization.