Problem.get_gradient

abstract Problem.get_gradient(*, decision, get_info=False, step=None)

Compute the gradient graph \(\nabla_y f\) for a given decision \(y\).

The gradient guides optimization algorithms such as gradient descent.

Parameters:

• decision (JaxGraph) – A decision graph at which to evaluate the gradient.

• get_info (bool) – Flag indicating if additional information should be returned for tracking purpose.

• step (int | None) – Training step number passed by the trainer. Useful for scheduling.

Returns:

A tuple containing: - Graph: The gradient graph with the same structure as decision. - dict: A dictionary of additional information (empty if get_info=False).

Raises:

NotImplementedError – If the subclass does not override this constructor.

Return type:

tuple[JaxGraph, dict]