Responsibilities

This page is a brain-dump of which parts of the framework are responsible for what.

The use-case should determine the input/output shape of the model. eg. For controlling the bot it should be state-in/controls-out. For physics prediction it might be (state,controller_inputs,dt)-in / state-out.

A trainer is a top-level entry point which ensures the model is created, converting and piping training-data to the model and saving the trained model.

runner.py/bot_manager.py are special-case consumers of the bot-controlling models.

The model does not have to be specialized for a use-case but it can be.

A model is responsible for defining the loss function and adjusting its weights when training data is provided.