Dear Lucy,

First of all, to apply deep learning algorithms to your problem, you'll need lots of data. By the few variables you described, maybe this isn't you case.

Generally, when choosing a machine learning algorithm, there are lots of questions about the data at hand and the considered application. For example: How many instances and classes there are? Do you need the training phase to be fast?

If you don't have at least a very complex problem, a lot of instances and structure to train a deep neural network, this shoudn't be your algorithm of choice.

In this case, I'd suggest you to play with simple algorithms first. For example, you could begin running a decision tree to see if some of your features provides a perfect separation between classes. Then you could move to linear algorithms and so on.

Edit: I'm leaving here the link of an empirical study comparing a number of supervised learning algorithms (https://www.cs.cornell.edu/~caruana/ctp/ct.papers/caruana.icml06.pdf)

Deep learning simply means that one suspects that the classification may not be linearly separable, that's all. That's why one would need multiple layers, that can, in turn be used to describe finer resolution of certain properties of the classification problem.

On the other hand this doesn't mean that any given deep learning network can solve any classification problem. The way to understand what's going on was described in this paper, almost thirty years ago:

http://lptms.u-psud.fr/membres/Mezard/Pdf/89_MN_JPA.pdf

I think as you have predefin features, you don’t need to use the deep learning algs. This is because they take a long processing time and a huge amount of data to ensure the right learnin.