Introduction to machine learning¶
MDODiscipline is costly to evaluate, it can be replaced by
SurrogateDiscipline cheap to evaluate, e.g. linear model, Kriging,
RBF regressor, …
SurrogateDiscipline is built from a few evaluations of this
MDODiscipline. This learning phase commonly relies on a regression
model calibrated by machine learning techniques. This is the reason why
GEMSEO provides a machine learning package which includes the
MLRegressionAlgo class implementing the concept of regression model.
In addition, this machine learning package has a much broader set of features
than regression: clustering, classification, dimension reduction, data scaling,
Machine learning algorithm baseclass¶
Machine learning is the art of building models from data, the latter being samples of properties of interest that can sometimes be sorted by group, such as inputs, outputs, categories, …
In the absence of such groups, the data can be analyzed through a study of commonalities, leading to plausible clusters. This is referred to as clustering, a branch of unsupervised learning dedicated to the detection of patterns in unlabeled data.
When data can be separated into at least two categories by a human, supervised learning can start with classification whose purpose is to model the relation between these categories and the properties of interest. Once trained, a classification model can predict the category corresponding to new property values.
When the distinction between inputs and outputs can be made among the data properties, another branch of supervised learning can be called: regression modeling. Once trained, a regression model can predict the outputs corresponding to new inputs values.
The quality of a machine learning algorithm can be measured
MLQualityMeasure either with respect to the learning dataset
or to a test dataset or using resampling methods, such as K-folds or
leave-one-out cross-validation techniques. The challenge is to avoid
over-learning the learning data leading to a loss of generality.
We often want to build models that are not too dataset-dependent. For that,
we want to maximize both a learning quality and a generalization quality.
In unsupervised learning, a quality measure can represent the robustness of
clusters definition while in supervised learning, a quality measure can be
interpreted as an error, whether it is a misclassification in the case of the
classification algorithms or a prediction one in the case of the regression
algorithms. This quality can often be improved by building machine learning
models from standardized data in such a way that the data properties have the
same order of magnitude.
Lastly, a machine learning algorithm often depends on hyperparameters to carefully tune in order to maximize the generalization power of the model.