Hyper parameter tuning for multiple Outputs using mxnet in R

Question

I currently try to build MLPs with multiple Outputs.

For single Output MLPs I normally use the H2o packge implementation which has a nice random grid search function implemented. Since H2o does not support multiple outputs I switched to the mxnet package.

Now I am trying to find a way to tune my parameters for my MLP. I can't find any package inside R which provides parameter tuning for multiple outputs and allows me to use mxnet.

Do you know any packages or do you have self implemented functions for a hyper parameter search? Thank you!

Edit in cause of comment:

With multiple outputs i mean multiple response variables (MIMO problem). For example one of my researched tasks is the prediction of a RRSB distribution. A RRSB distribution has two parameters: n,x. I hope this clears your question

Can you be more specific? What do you mean by multiple outputs? — acylam, Sep 26 '17 at 13:27
I tried to be and updated my post. I hope this answers your question — Paul, Sep 26 '17 at 16:25

score 1 · Answer 1 · answered Oct 04 '17 at 04:39

Have you consider Caret? I think it might do it as it provides both hyper-parameters tuning https://topepo.github.io/caret/model-training-and-tuning.html and support some mxnet models, but unsure whether a MLP with multiple outputs is supported.

I personally use a grid or random search approach simply by sampling the required parameters from list of possible values associated to each. It involves looping over different hyper-parameter selection and keeping a log of each configuration performance.

I considered Caret. It is a good package but it doesn't support multiple outputs. I think I will also have to use grid random search. I just would have preffered a package. But thx for the input — Paul, Oct 04 '17 at 12:13

score 0 · Answer 2 · answered Mar 07 '18 at 01:18

This tutorial explains how you can build a Multi-Layer Perceptron using MXNet, in that particular case which has a Softmax layer as a final layer to classify the input as one of 10 different classes (MNIST problem).

In your case I believe what you want is a MLP with a final fully connected layer of 2 units, one for each variable you are trying to predict. Look here for an example of linear regression. Your loss function would need to be adapted to your specific problem.

Hyper parameter tuning for multiple Outputs using mxnet in R

2 Answers2