src.variants.metamodels.neural_networks module

This module is responsible for training and inference of neural networks.

It includes the two functions train_mlp_model and infer_mlp_model for training and inference of multi layer perceptrons. In addition, it offers a custom loss function to learn an upper bound. See details in their own documentations.

It is a new implementation of the research we have done in student thesis [1].

Thesis: [1] J. Wang, „Maschinelles Lernen zur Metamodellierung von Fehlern in der Simulation automatisierter Fahrzeuge,“ Master’s Thesis, Technical University of Munich, Munich, Germany, 2020.

Contact person: Stefan Riedmaier Creation date: 04.06.2020 Python version: 3.8

src.variants.metamodels.neural_networks.infer_mlp_model(x, model)

This function predicts outputs of a multi-layer perceptron model for inputs x.

Parameters:

• x (np.ndarray) – test data set inputs

• model – multi-layer perceptron

Returns:

predicted outputs on the test data set

Return type:

np.ndarray

src.variants.metamodels.neural_networks.train_mlp_model(x, y)

This function trains a multi-layer perceptron (MLP).

Parameters:

• x (np.ndarray) – training input data

• y (np.ndarray) – training output data

Returns:

trained multi-layer perceptron

src.variants.metamodels.neural_networks.upper_bound_loss(y_true, y_pred)

This function defines a loss calculation to learn an upper bound on data.

Parameters:

• y_true (np.ndarray) – ground truth outputs

• y_pred (np.ndarray) – predicted outputs

Returns:

loss value

Return type:

float