Defining an optimization element

In the optimization module, the objectives and constraints are both optimization elements (Objective inherits from Constraint). They are both callable and have a function attribute which defines the function to be optimized or to comply with.

There are two ways to go from a raw variable provided by the user to an objective value:

• For a trained model (left in diagram below), the user can provide a DataProcessor to process the raw variable in the same way as the training data, and a model (implementing torch.nn.Module) to compute the objective value.

• Otherwise (right below), the user can provide a function that takes the raw variable as input and returns the objective value. This function is then wrapped in a Objective object. In that case, the signature of the functino should fit the UdaoFunction signature, with two parameters: input_variables and input_parameters.

In the case a DataProcessor is provided, it can be taken into account in two ways:

• At the problem level, being a preprocessing step for all the objectives and constraints. In that case, the DataProcessor is provided to the BaseProblem object.

• At the objective level, being a preprocessing step for this objective only. In that case, the DataProcessor is bundled with the model in a ModelComponent object. This is useful when some functions require a data processor but others do not. However, it is not possible to use this method for gradient descent methods, as will be explained in The case of MOGD.