# Create a general linear model metamodel¶

In this example we are going to create a global approximation of a model response using a general linear model. We show how to use the GeneralLinearModelAlgorithm class, which estimates the parameters of the model.

[1]:

from __future__ import print_function
import openturns as ot


We create a model and a sample from this model.

[2]:

ot.RandomGenerator.SetSeed(0)
dimension = 2
input_names = ['x1', 'x2']
formulas = ['cos(x1 + x2)']
model = ot.SymbolicFunction(input_names, formulas)
distribution = ot.Normal(dimension)
x  = distribution.getSample(100)
y = model(x)


We create a GeneralLinearModelAlgorithm based on a linear basis. The run method estimats the coefficients of the trend and the hyperparameters of the covariance model.

[3]:

basis = ot.LinearBasisFactory(dimension).build()
covarianceModel = ot.SquaredExponential([1]*dimension, [1.0])
algo = ot.GeneralLinearModelAlgorithm(x, y, covarianceModel, basis)
algo.run()
result = algo.getResult()


We see that the trend coefficients have been estimated.

[4]:

result.getTrendCoefficients()

[4]:

[class=Point name=Unnamed dimension=3 values=[-0.111515,-0.0142378,-0.0129734]]


The parameters of the covariance models also have been estimated.

[5]:

result.getCovarianceModel()

[5]:


SquaredExponential(scale=[0.810148,1.52486], amplitude=[0.404099])

The getMetaModel method returns the metamodel where the parameters have been estimated.

[6]:

responseSurface = result.getMetaModel()


Plot the output of our model depending on x2 with x1=0.5.

[7]:

xmin = -5.
xmax = 5.
x1value = 0.5
parametricModelGraph = ot.ParametricFunction(model, [0], [x1value]).draw(xmin,xmax)
parametricModelGraph.setColors(["red"])
graphMetamodel = ot.ParametricFunction(responseSurface, [0], [x1value]).draw(xmin,xmax)
graphMetamodel.setColors(["blue"])

[7]:

This GeneralLinearModelAlgorithm class is the main building block of the KrigingAlgorithm. This is why most basic use cases are based on the KrigingAlgorithm instead of the GeneralLinearModelAlgorithm, because this allows to condition the gaussian process.