Polynomial chaos expansion (PCE)

A PCERegressor is a PCE model based on OpenTURNS.

from __future__ import annotations

from matplotlib import pyplot as plt
from numpy import array

from gemseo import configure_logger
from gemseo import create_discipline
from gemseo import create_parameter_space
from gemseo import sample_disciplines
from gemseo.mlearning import create_regression_model

configure_logger()

<RootLogger root (INFO)>

Problem

In this example, we represent the function \(f(x)=(6x-2)^2\sin(12x-4)\) [FSK08] by the AnalyticDiscipline

discipline = create_discipline(
    "AnalyticDiscipline",
    name="f",
    expressions={"y": "(6*x-2)**2*sin(12*x-4)"},
)

and seek to approximate it over the input space

input_space = create_parameter_space()
input_space.add_random_variable("x", "OTUniformDistribution")

To do this, we create a training dataset with 6 equispaced points:

training_dataset = sample_disciplines(
    [discipline], input_space, "y", algo_name="PYDOE_FULLFACT", n_samples=10
)

WARNING - 20:35:11: No coupling in MDA, switching chain_linearize to True.
   INFO - 20:35:11: *** Start Sampling execution ***
   INFO - 20:35:11: Sampling
   INFO - 20:35:11:    Disciplines: f
   INFO - 20:35:11:    MDO formulation: MDF
   INFO - 20:35:11: Running the algorithm PYDOE_FULLFACT:
   INFO - 20:35:11:     10%|█         | 1/10 [00:00<00:00, 661.67 it/sec]
   INFO - 20:35:11:     20%|██        | 2/10 [00:00<00:00, 1086.47 it/sec]
   INFO - 20:35:11:     30%|███       | 3/10 [00:00<00:00, 1423.25 it/sec]
   INFO - 20:35:11:     40%|████      | 4/10 [00:00<00:00, 1711.96 it/sec]
   INFO - 20:35:11:     50%|█████     | 5/10 [00:00<00:00, 1960.69 it/sec]
   INFO - 20:35:11:     60%|██████    | 6/10 [00:00<00:00, 2174.34 it/sec]
   INFO - 20:35:11:     70%|███████   | 7/10 [00:00<00:00, 2360.90 it/sec]
   INFO - 20:35:11:     80%|████████  | 8/10 [00:00<00:00, 2505.00 it/sec]
   INFO - 20:35:11:     90%|█████████ | 9/10 [00:00<00:00, 2648.85 it/sec]
   INFO - 20:35:11:    100%|██████████| 10/10 [00:00<00:00, 2780.45 it/sec]
   INFO - 20:35:11: *** End Sampling execution ***

Basics

Training

Then, we train an PCE regression model from these samples:

model = create_regression_model("PCERegressor", training_dataset)
model.learn()

WARNING - 20:35:11: Remove input data transformation because PCERegressor does not support transformers.

Prediction

Once it is built, we can predict the output value of \(f\) at a new input point:

input_value = {"x": array([0.65])}
output_value = model.predict(input_value)
output_value

{'y': array([-0.81106394])}

as well as its Jacobian value:

jacobian_value = model.predict_jacobian(input_value)
jacobian_value

{'y': {'x': array([[18.2279622]])}}

Plotting

Of course, you can see that the quadratic model is no good at all here:

test_dataset = sample_disciplines(
    [discipline], input_space, "y", algo_name="PYDOE_FULLFACT", n_samples=100
)
input_data = test_dataset.get_view(variable_names=model.input_names).to_numpy()
reference_output_data = test_dataset.get_view(variable_names="y").to_numpy().ravel()
predicted_output_data = model.predict(input_data).ravel()
plt.plot(input_data.ravel(), reference_output_data, label="Reference")
plt.plot(input_data.ravel(), predicted_output_data, label="Regression - Basics")
plt.grid()
plt.legend()
plt.show()

WARNING - 20:35:11: No coupling in MDA, switching chain_linearize to True.
   INFO - 20:35:11: *** Start Sampling execution ***
   INFO - 20:35:11: Sampling
   INFO - 20:35:11:    Disciplines: f
   INFO - 20:35:11:    MDO formulation: MDF
   INFO - 20:35:11: Running the algorithm PYDOE_FULLFACT:
   INFO - 20:35:11:      1%|          | 1/100 [00:00<00:00, 3446.43 it/sec]
   INFO - 20:35:11:      2%|▏         | 2/100 [00:00<00:00, 3425.32 it/sec]
   INFO - 20:35:11:      3%|▎         | 3/100 [00:00<00:00, 3680.29 it/sec]
   INFO - 20:35:11:      4%|▍         | 4/100 [00:00<00:00, 3884.51 it/sec]
   INFO - 20:35:11:      5%|▌         | 5/100 [00:00<00:00, 4028.34 it/sec]
   INFO - 20:35:11:      6%|▌         | 6/100 [00:00<00:00, 4060.31 it/sec]
   INFO - 20:35:11:      7%|▋         | 7/100 [00:00<00:00, 4137.56 it/sec]
   INFO - 20:35:11:      8%|▊         | 8/100 [00:00<00:00, 4220.68 it/sec]
   INFO - 20:35:11:      9%|▉         | 9/100 [00:00<00:00, 4281.36 it/sec]
   INFO - 20:35:11:     10%|█         | 10/100 [00:00<00:00, 4346.43 it/sec]
   INFO - 20:35:11:     11%|█         | 11/100 [00:00<00:00, 4365.76 it/sec]
   INFO - 20:35:11:     12%|█▏        | 12/100 [00:00<00:00, 4407.71 it/sec]
   INFO - 20:35:11:     13%|█▎        | 13/100 [00:00<00:00, 4450.37 it/sec]
   INFO - 20:35:11:     14%|█▍        | 14/100 [00:00<00:00, 4488.97 it/sec]
   INFO - 20:35:11:     15%|█▌        | 15/100 [00:00<00:00, 4516.16 it/sec]
   INFO - 20:35:11:     16%|█▌        | 16/100 [00:00<00:00, 4521.55 it/sec]
   INFO - 20:35:11:     17%|█▋        | 17/100 [00:00<00:00, 4550.01 it/sec]
   INFO - 20:35:11:     18%|█▊        | 18/100 [00:00<00:00, 4576.44 it/sec]
   INFO - 20:35:11:     19%|█▉        | 19/100 [00:00<00:00, 4604.33 it/sec]
   INFO - 20:35:11:     20%|██        | 20/100 [00:00<00:00, 4631.78 it/sec]
   INFO - 20:35:11:     21%|██        | 21/100 [00:00<00:00, 4616.86 it/sec]
   INFO - 20:35:11:     22%|██▏       | 22/100 [00:00<00:00, 4630.64 it/sec]
   INFO - 20:35:11:     23%|██▎       | 23/100 [00:00<00:00, 4649.33 it/sec]
   INFO - 20:35:11:     24%|██▍       | 24/100 [00:00<00:00, 4668.12 it/sec]
   INFO - 20:35:11:     25%|██▌       | 25/100 [00:00<00:00, 4686.37 it/sec]
   INFO - 20:35:11:     26%|██▌       | 26/100 [00:00<00:00, 4669.32 it/sec]
   INFO - 20:35:11:     27%|██▋       | 27/100 [00:00<00:00, 4678.44 it/sec]
   INFO - 20:35:11:     28%|██▊       | 28/100 [00:00<00:00, 4691.43 it/sec]
   INFO - 20:35:11:     29%|██▉       | 29/100 [00:00<00:00, 4705.41 it/sec]
   INFO - 20:35:11:     30%|███       | 30/100 [00:00<00:00, 4720.48 it/sec]
   INFO - 20:35:11:     31%|███       | 31/100 [00:00<00:00, 4714.24 it/sec]
   INFO - 20:35:11:     32%|███▏      | 32/100 [00:00<00:00, 4722.82 it/sec]
   INFO - 20:35:11:     33%|███▎      | 33/100 [00:00<00:00, 4734.30 it/sec]
   INFO - 20:35:11:     34%|███▍      | 34/100 [00:00<00:00, 4743.43 it/sec]
   INFO - 20:35:11:     35%|███▌      | 35/100 [00:00<00:00, 4751.29 it/sec]
   INFO - 20:35:11:     36%|███▌      | 36/100 [00:00<00:00, 4745.58 it/sec]
   INFO - 20:35:11:     37%|███▋      | 37/100 [00:00<00:00, 4755.01 it/sec]
   INFO - 20:35:11:     38%|███▊      | 38/100 [00:00<00:00, 4765.54 it/sec]
   INFO - 20:35:11:     39%|███▉      | 39/100 [00:00<00:00, 4775.58 it/sec]
   INFO - 20:35:11:     40%|████      | 40/100 [00:00<00:00, 4785.97 it/sec]
   INFO - 20:35:11:     41%|████      | 41/100 [00:00<00:00, 4780.70 it/sec]
   INFO - 20:35:11:     42%|████▏     | 42/100 [00:00<00:00, 4788.02 it/sec]
   INFO - 20:35:11:     43%|████▎     | 43/100 [00:00<00:00, 4797.44 it/sec]
   INFO - 20:35:11:     44%|████▍     | 44/100 [00:00<00:00, 4805.60 it/sec]
   INFO - 20:35:11:     45%|████▌     | 45/100 [00:00<00:00, 4813.42 it/sec]
   INFO - 20:35:11:     46%|████▌     | 46/100 [00:00<00:00, 4804.95 it/sec]
   INFO - 20:35:11:     47%|████▋     | 47/100 [00:00<00:00, 4811.98 it/sec]
   INFO - 20:35:11:     48%|████▊     | 48/100 [00:00<00:00, 4820.12 it/sec]
   INFO - 20:35:11:     49%|████▉     | 49/100 [00:00<00:00, 4826.70 it/sec]
   INFO - 20:35:11:     50%|█████     | 50/100 [00:00<00:00, 4832.93 it/sec]
   INFO - 20:35:11:     51%|█████     | 51/100 [00:00<00:00, 4827.35 it/sec]
   INFO - 20:35:11:     52%|█████▏    | 52/100 [00:00<00:00, 4832.68 it/sec]
   INFO - 20:35:11:     53%|█████▎    | 53/100 [00:00<00:00, 4786.78 it/sec]
   INFO - 20:35:11:     54%|█████▍    | 54/100 [00:00<00:00, 4789.64 it/sec]
   INFO - 20:35:11:     55%|█████▌    | 55/100 [00:00<00:00, 4786.53 it/sec]
   INFO - 20:35:11:     56%|█████▌    | 56/100 [00:00<00:00, 4790.07 it/sec]
   INFO - 20:35:11:     57%|█████▋    | 57/100 [00:00<00:00, 4793.97 it/sec]
   INFO - 20:35:11:     58%|█████▊    | 58/100 [00:00<00:00, 4799.83 it/sec]
   INFO - 20:35:11:     59%|█████▉    | 59/100 [00:00<00:00, 4806.15 it/sec]
   INFO - 20:35:11:     60%|██████    | 60/100 [00:00<00:00, 4805.94 it/sec]
   INFO - 20:35:11:     61%|██████    | 61/100 [00:00<00:00, 4807.90 it/sec]
   INFO - 20:35:11:     62%|██████▏   | 62/100 [00:00<00:00, 4811.49 it/sec]
   INFO - 20:35:11:     63%|██████▎   | 63/100 [00:00<00:00, 4816.29 it/sec]
   INFO - 20:35:11:     64%|██████▍   | 64/100 [00:00<00:00, 4820.69 it/sec]
   INFO - 20:35:11:     65%|██████▌   | 65/100 [00:00<00:00, 4816.78 it/sec]
   INFO - 20:35:11:     66%|██████▌   | 66/100 [00:00<00:00, 4819.78 it/sec]
   INFO - 20:35:11:     67%|██████▋   | 67/100 [00:00<00:00, 4825.26 it/sec]
   INFO - 20:35:11:     68%|██████▊   | 68/100 [00:00<00:00, 4831.33 it/sec]
   INFO - 20:35:11:     69%|██████▉   | 69/100 [00:00<00:00, 4837.15 it/sec]
   INFO - 20:35:11:     70%|███████   | 70/100 [00:00<00:00, 4837.00 it/sec]
   INFO - 20:35:11:     71%|███████   | 71/100 [00:00<00:00, 4838.98 it/sec]
   INFO - 20:35:11:     72%|███████▏  | 72/100 [00:00<00:00, 4840.75 it/sec]
   INFO - 20:35:11:     73%|███████▎  | 73/100 [00:00<00:00, 4844.15 it/sec]
   INFO - 20:35:11:     74%|███████▍  | 74/100 [00:00<00:00, 4848.91 it/sec]
   INFO - 20:35:11:     75%|███████▌  | 75/100 [00:00<00:00, 4847.41 it/sec]
   INFO - 20:35:11:     76%|███████▌  | 76/100 [00:00<00:00, 4848.91 it/sec]
   INFO - 20:35:11:     77%|███████▋  | 77/100 [00:00<00:00, 4852.48 it/sec]
   INFO - 20:35:11:     78%|███████▊  | 78/100 [00:00<00:00, 4857.11 it/sec]
   INFO - 20:35:11:     79%|███████▉  | 79/100 [00:00<00:00, 4861.14 it/sec]
   INFO - 20:35:11:     80%|████████  | 80/100 [00:00<00:00, 4860.21 it/sec]
   INFO - 20:35:11:     81%|████████  | 81/100 [00:00<00:00, 4862.02 it/sec]
   INFO - 20:35:11:     82%|████████▏ | 82/100 [00:00<00:00, 4864.82 it/sec]
   INFO - 20:35:11:     83%|████████▎ | 83/100 [00:00<00:00, 4867.75 it/sec]
   INFO - 20:35:11:     84%|████████▍ | 84/100 [00:00<00:00, 4871.23 it/sec]
   INFO - 20:35:11:     85%|████████▌ | 85/100 [00:00<00:00, 4867.77 it/sec]
   INFO - 20:35:11:     86%|████████▌ | 86/100 [00:00<00:00, 4868.47 it/sec]
   INFO - 20:35:11:     87%|████████▋ | 87/100 [00:00<00:00, 4870.98 it/sec]
   INFO - 20:35:11:     88%|████████▊ | 88/100 [00:00<00:00, 4873.88 it/sec]
   INFO - 20:35:11:     89%|████████▉ | 89/100 [00:00<00:00, 4876.84 it/sec]
   INFO - 20:35:11:     90%|█████████ | 90/100 [00:00<00:00, 4875.65 it/sec]
   INFO - 20:35:11:     91%|█████████ | 91/100 [00:00<00:00, 4876.29 it/sec]
   INFO - 20:35:11:     92%|█████████▏| 92/100 [00:00<00:00, 4879.38 it/sec]
   INFO - 20:35:11:     93%|█████████▎| 93/100 [00:00<00:00, 4882.29 it/sec]
   INFO - 20:35:11:     94%|█████████▍| 94/100 [00:00<00:00, 4885.20 it/sec]
   INFO - 20:35:11:     95%|█████████▌| 95/100 [00:00<00:00, 4883.49 it/sec]
   INFO - 20:35:11:     96%|█████████▌| 96/100 [00:00<00:00, 4885.38 it/sec]
   INFO - 20:35:11:     97%|█████████▋| 97/100 [00:00<00:00, 4889.11 it/sec]
   INFO - 20:35:11:     98%|█████████▊| 98/100 [00:00<00:00, 4892.54 it/sec]
   INFO - 20:35:11:     99%|█████████▉| 99/100 [00:00<00:00, 4895.84 it/sec]
   INFO - 20:35:11:    100%|██████████| 100/100 [00:00<00:00, 4895.31 it/sec]
   INFO - 20:35:11: *** End Sampling execution ***

Settings

The PCERegressor has many options defined in the PCERegressor_Settings Pydantic model.

Degree

model = create_regression_model("PCERegressor", training_dataset, degree=3)
model.learn()

WARNING - 20:35:11: Remove input data transformation because PCERegressor does not support transformers.

and see that this model seems to be better:

predicted_output_data_ = model.predict(input_data).ravel()
plt.plot(input_data.ravel(), reference_output_data, label="Reference")
plt.plot(input_data.ravel(), predicted_output_data, label="Regression - Basics")
plt.plot(input_data.ravel(), predicted_output_data_, label="Regression - Degree(3)")
plt.grid()
plt.legend()
plt.show()