Tracking v3

[andyk]

Copying some thoughts about the next version of Tracking from #178.

• Tracking functionality can be separated into components at two different conceptual "layers":
  • A lower more general/flexible layer that is not specific to agent development, which just exposes MLflow-like tracking semantics to the developer, i.e., log_parameter, log_metric, log_model, log_artifacts.
  • A higher layer that provides agent-aware semantics. Essentially, what you have proposed in BaseTracker -- it is aware of evaluation, learning, rollouts/episodes
• Also, would it make sense for the runtime to be aware of the tracking module as a special component? I'm not sure I like that better than the approach you're using here. The way you got around this in the current prototype is by having a generic dictionary get attached to every component in a Component's dependency DAG at the Component's creation time. And then the Tracking component accesses that information to log the component_name and entry_point. This is the main way the runtime has to communicate to components/classes it manages (i.e., instantiates)