Need documentation on what is allowed in training files including:
-Overview of the different levels of attributes: global, training_et, agent, problem_set
-What attributes are legal at each level
-What attributes are open-ended (for example depends on agent or outerloop)
-Documentation of eval functions