PCE regression

We want to approximate a discipline with two inputs and two outputs:

• \(y_1=1+2x_1+3x_2\)

• \(y_2=-1-2x_1-3x_2\)

over the unit hypercube \([0,1]\times[0,1]\).

from __future__ import annotations

from gemseo.api import configure_logger
from gemseo.api import create_design_space
from gemseo.api import create_discipline
from gemseo.api import create_parameter_space
from gemseo.api import create_scenario
from gemseo.mlearning.api import create_regression_model
from gemseo.mlearning.api import import_regression_model
from numpy import array

configure_logger()

<RootLogger root (INFO)>

Create the discipline to learn

We can implement this analytic discipline by means of the AnalyticDiscipline class.

expressions = {"y_1": "1+2*x_1+3*x_2", "y_2": "-1-2*x_1-3*x_2"}
discipline = create_discipline(
    "AnalyticDiscipline", name="func", expressions=expressions
)

Create the input sampling space

We create the input sampling space by adding the variables one by one.

design_space = create_design_space()
design_space.add_variable("x_1", l_b=0.0, u_b=1.0)
design_space.add_variable("x_2", l_b=0.0, u_b=1.0)

Create the learning set

We can build a learning set by means of a DOEScenario with a full factorial design of experiments. The number of samples can be equal to 9 for example.

scenario = create_scenario(
    [discipline], "DisciplinaryOpt", "y_1", design_space, scenario_type="DOE"
)
scenario.execute({"algo": "fullfact", "n_samples": 9})

    INFO - 13:29:46:
    INFO - 13:29:46: *** Start DOEScenario execution ***
    INFO - 13:29:46: DOEScenario
    INFO - 13:29:46:    Disciplines: func
    INFO - 13:29:46:    MDO formulation: DisciplinaryOpt
    INFO - 13:29:46: Optimization problem:
    INFO - 13:29:46:    minimize y_1(x_1, x_2)
    INFO - 13:29:46:    with respect to x_1, x_2
    INFO - 13:29:46:    over the design space:
    INFO - 13:29:46:    +------+-------------+-------+-------------+-------+
    INFO - 13:29:46:    | name | lower_bound | value | upper_bound | type  |
    INFO - 13:29:46:    +------+-------------+-------+-------------+-------+
    INFO - 13:29:46:    | x_1  |      0      |  None |      1      | float |
    INFO - 13:29:46:    | x_2  |      0      |  None |      1      | float |
    INFO - 13:29:46:    +------+-------------+-------+-------------+-------+
    INFO - 13:29:46: Solving optimization problem with algorithm fullfact:
    INFO - 13:29:46: ...   0%|          | 0/9 [00:00<?, ?it]
    INFO - 13:29:46: ...  11%|█         | 1/9 [00:00<00:00, 343.06 it/sec, obj=1]
    INFO - 13:29:46: ...  22%|██▏       | 2/9 [00:00<00:00, 553.19 it/sec, obj=2]
    INFO - 13:29:46: ...  33%|███▎      | 3/9 [00:00<00:00, 699.71 it/sec, obj=3]
    INFO - 13:29:46: ...  44%|████▍     | 4/9 [00:00<00:00, 810.38 it/sec, obj=2.5]
    INFO - 13:29:46: ...  56%|█████▌    | 5/9 [00:00<00:00, 896.37 it/sec, obj=3.5]
    INFO - 13:29:46: ...  67%|██████▋   | 6/9 [00:00<00:00, 963.47 it/sec, obj=4.5]
    INFO - 13:29:46: ...  78%|███████▊  | 7/9 [00:00<00:00, 1025.04 it/sec, obj=4]
    INFO - 13:29:46: ...  89%|████████▉ | 8/9 [00:00<00:00, 1039.90 it/sec, obj=5]
    INFO - 13:29:46: ... 100%|██████████| 9/9 [00:00<00:00, 1079.55 it/sec, obj=6]
    INFO - 13:29:46: Optimization result:
    INFO - 13:29:46:    Optimizer info:
    INFO - 13:29:46:       Status: None
    INFO - 13:29:46:       Message: None
    INFO - 13:29:46:       Number of calls to the objective function by the optimizer: 9
    INFO - 13:29:46:    Solution:
    INFO - 13:29:46:       Objective: 1.0
    INFO - 13:29:46:       Design space:
    INFO - 13:29:46:       +------+-------------+-------+-------------+-------+
    INFO - 13:29:46:       | name | lower_bound | value | upper_bound | type  |
    INFO - 13:29:46:       +------+-------------+-------+-------------+-------+
    INFO - 13:29:46:       | x_1  |      0      |   0   |      1      | float |
    INFO - 13:29:46:       | x_2  |      0      |   0   |      1      | float |
    INFO - 13:29:46:       +------+-------------+-------+-------------+-------+
    INFO - 13:29:46: *** End DOEScenario execution (time: 0:00:00.017357) ***

{'eval_jac': False, 'n_samples': 9, 'algo': 'fullfact'}

Create the regression model

Then, we build the linear regression model from the database and displays this model.

prob_space = create_parameter_space()
prob_space.add_random_variable("x_1", "OTUniformDistribution")
prob_space.add_random_variable("x_2", "OTUniformDistribution")
dataset = scenario.export_to_dataset(opt_naming=False)
model = create_regression_model(
    "PCERegressor", data=dataset, probability_space=prob_space, transformer=None
)
model.learn()
print(model)

PCERegressor(cleaning_options=CleaningOptions(max_considered_terms=100, most_significant=20, significance_factor=0.0001), degree=2, hyperbolic_parameter=1.0, n_quadrature_points=0, probability_space=+----------------------------------------------------------------------------------+
|                                 Parameter space                                  |
+------+-------------+-------+-------------+-------+-------------------------------+
| name | lower_bound | value | upper_bound | type  |      Initial distribution     |
+------+-------------+-------+-------------+-------+-------------------------------+
| x_1  |      0      |  0.5  |      1      | float | Uniform(lower=0.0, upper=1.0) |
| x_2  |      0      |  0.5  |      1      | float | Uniform(lower=0.0, upper=1.0) |
+------+-------------+-------+-------------+-------+-------------------------------+, use_cleaning=False, use_lars=False)
   based on the OpenTURNS library
   built from 9 learning samples

Predict output

Once it is built, we can use it for prediction.

input_value = {"x_1": array([1.0]), "x_2": array([2.0])}
output_value = model.predict(input_value)
print(output_value)

{'y_1': array([9.])}

Save the regression model

Lastly, we save the model.

directory = model.save()

Load the regression model

In an other study, we could load this model.

loaded_model = import_regression_model(directory)
print(loaded_model)

PCERegressor(cleaning_options=CleaningOptions(max_considered_terms=100, most_significant=20, significance_factor=0.0001), degree=2, hyperbolic_parameter=1.0, n_quadrature_points=0, probability_space=+----------------------------------------------------------------------------------+
|                                 Parameter space                                  |
+------+-------------+-------+-------------+-------+-------------------------------+
| name | lower_bound | value | upper_bound | type  |      Initial distribution     |
+------+-------------+-------+-------------+-------+-------------------------------+
| x_1  |      0      |  0.5  |      1      | float | Uniform(lower=0.0, upper=1.0) |
| x_2  |      0      |  0.5  |      1      | float | Uniform(lower=0.0, upper=1.0) |
+------+-------------+-------+-------------+-------+-------------------------------+, use_cleaning=False, use_lars=False)
   based on the OpenTURNS library
   built from 0 learning samples

Use the loaded regression model

And use it!

print(loaded_model.predict(input_value))

{'y_1': array([9.])}