[learning_to_simulate] Clarification about training: simulator() vs. get_predicted_and_target_normalized_accelerations()

Hello,

Thank you for sharing this great project and code. I have a question regarding the call to both the `get_predicted_and_target_normalized_accelerations()` and `simulator()` functions within the `one_step_estimator_fn()`, which defines the main training loop logic.

Is the former function `simulator()` strictly included in the training loop for evaluation/metrics purposes? i.e. its position prediction and corresponding gradients do not influence the optimization results at all -- is this understanding correct?

On a related note, since I am not highly familiar with TF-v1 -- when `simulator()` is called, can you confirm whether or not this specific operation is included in the computation graph? I would assume it should not be, since its gradients shouldn't be required for optimization, but am unsure of whether this happens by virtue of having no connection with the optimizer.

[Referencing code here.](https://github.com/deepmind/deepmind-research/blob/master/learning_to_simulate/train.py#L361)

Thank you very much for your help.