Optimized DNN parameter

Dear experts,
I wonder if there is a recommendation to optimise the DNN parameters. Indeed I started from the tutorial parameters and I vary separatly other parameters. But they may be correlated (maybe I should vary 2 at the same time to see improvement…), so I wonder if there is a recommended way to do that?


I don’t have any concrete tips for choosing the hyperparamters, but having a look at is currently considered best practises can be informative. E.g. under the subsection “Choosing Hyperparamters” in this post.

The article is a summary of the 2016 book “Deep Learning” and states that the learning rate is important to choose correctly initially. Other parameters are then best found doing a grid search or a random search. Doing an automatic search in TMVA for the DNN is currently not possible, but a limited manual search is feasible.