#35 Implement TrainingMethod SARPROP in MethodMLP

Dear TMVA developers,

I have used TMVA and its neural network implementation MLP for my master's thesis on tau lepton identification (still to be published). Over the course of my thesis, I have studied the SARPROP training algorithm and its effects on classifier performance. The algorithm is described here and is based on iRProp–, a popular modification of classical gradient-based approaches.

I have found that (at least for my classification problem) SARPROP resulted in a classifier performance comparable to BFGS. At the same time, it was approximately twice as fast due to the reduced computational complexity. (No need for matrix inversion, no LineSearch algorithm, etc.)

Due to these advantages, I'd like to implement SAPROP in TMVA. I've prepared a GitHub fork of ROOT 5.34/34 where I've added an implementation to the class MethodMLP. You can review the relevant changes in this commit. (It's admittedly very hacky so far.)

I hope this is the correct place to discuss this addition. Please let me know if the mailing list would be more appropriate.

Regards,

Nico Madysa