Refactor the way learning/optimization methods are initialized

[fwilhelmi]

At this moment, all the learning/optimization methods are treated equally, thus providing the same (or similar) methods for initializing variables, keeping track of information, etc.

Since different learning/optimization methods may significantly differ among them, we propose to refactor the way those methods are initialized, so that the "agent" component is agnostic of such a procedure. In particular, we propose to provide different input files (e.g., json), which are specific for each learning/optimization approach. Accordingly, each file will be parsed differently, thus initializing the necessary variables within the code to carry out the optimization operation.