blankjul / pysurrogate

pysurrogate - Surrogate Models

The aim of this project is to build an interface to find the best surrogate for observations made on a blackbox function. The nature of a blackbox function do not let us make any assumptions about the data and therefore the model selection needs to be done more carefully.

Installation

pip install pysurrogate

Surrogates

Currently the following surrogate models are implemented:

• DACE (Matlab Version, matlab library for python needs to be installed)
• RBF (Scipy)
• NN (pytorch)

Usage

The method that should be used for interfacing is the pysurrogate.optimize.fit and pysurrogate.optimize.predict method.

import numpy as np
import matplotlib.pyplot as plt
from pysurrogate.optimize import fit, predict


if __name__ == '__main__':
    # number of samples we will use for this example
    n_samples = 20

    # ---------------------------------------------------------
    # Example 1: One input variable and one target
    # ---------------------------------------------------------

    X = np.random.rand(n_samples, 1) * 4 * np.pi
    Y = np.cos(X)

    # fit the model and predict the data
    model = fit(X, Y, n_folds=3, disp=True, normalize_X=True, normalize_Y=True)
    _X = np.linspace(0, 4 * np.pi, 1000)
    _Y = predict(model, _X)

    plt.scatter(X, Y, label="Observations")
    plt.plot(_X, _Y, label="True")
    plt.show()

    # ---------------------------------------------------------
    # Example 2: Two input variables and two targets.
    #            Normalize before building the model and use only an RBF implementation with a specific kernel
    #            Finally validate the model error on the true function.
    # ---------------------------------------------------------

    X = (np.random.rand(n_samples, 2) * 200) + 500
    func_eval = lambda X: np.concatenate([np.sum(np.square(X), axis=1)[:, None], np.sum(np.sqrt(X), axis=1)[:, None]], axis=1)
    Y = func_eval(X)

    # fit the model and predict the data
    model = fit(X, Y, n_folds=3, disp=True, normalize_X=True, normalize_Y=True)

    # create two dimensional data to test the
    M = np.meshgrid(np.linspace(100, 200, 1000), np.linspace(100, 200, 1000))
    _X = np.concatenate([X[:, :, None] for e in X], axis=2).reshape(n_samples * n_samples, 2)
    _Y = predict(model, _X)
    print(np.mean(np.abs(_Y - func_eval(_X)), axis=0))

Contact

Feel free to contact me if you have any question:

Julian Blank (blankjul [at] egr.msu.edu)

Michigan State University

Computational Optimization and Innovation Laboratory (COIN)

East Lansing, MI 48824, USA