Figuring out a good Neural Network hidden layer structure for this dataset structure

[tlg265]

I'm working on a Regression problem where I'm going to use Neural Networks for predictions. I have already preprocessed the input dataset.

After the preprocessing I have approximately the following (listing just the key points):

• 10 continuous variables
• 20 discrete variables which are converted to numeric with one hot encoding

After the one hot encoding I have:

• 300 numeric variables / columns (the one hot encoding created a new column for each discrete value)

I know that when training Neural Networks you have to try multiple combinatioins, but I want to open a debate about this.

My Question: Based on the key points above, what would be a good starting point for the hidden layer structure of the Neural Networks?

Thanks!