class_path_and_name

Specify the module and class names of the python surrogate

Specification

• Alias: None

• Arguments: STRING

Description

Specify the module and class names of the python surrogate. At minimum, the class must contain the methods ‘construct’ and ‘predict’. Dakota passes to the former a matrix (num samples x num variables) of variable values and a matrix (num samples x num components) of responses to use to train the surrogate, where num components is 1 for scalar surrogates. The predict method receives a matrix (num samples x num variables) and expects a matrix (num samples x num components) of predicted values to be returned. Optional ‘gradient’ and ‘hessian’ methods are supported as well.

Examples

Use of python-based surrogates in Dakota is enabled with the model block via:

model,
    id_model = 'SURR'
    surrogate global,
    dace_method_pointer = 'DACE'
    experimental_python
        class_path_and_name = "surrogate_polynomial.Surrogate"

An example python class that implements a linear regression surrogate that works for both scalar and vector-valued responses and that includes gradient and (trivial) hessian values is the following:

class Surrogate:

    def __init__(self):
        self.coeffs = None

    def construct(self, var, resp):
        var2 = np.hstack((np.ones((var.shape[0], 1)), var))
        self.coeffs = np.zeros(var.shape[1])
        z = np.linalg.inv(np.dot(var2.T, var2))
        self.coeffs = np.dot(z, np.dot(var2.T, resp))
        return

    def predict(self, pts):
        return self.coeffs[0]+pts.dot(self.coeffs[1:])

    def gradient(self, pts):
        grad = self.coeffs[1:]
        return grad.T

    def hessian(self, pts):
        num_vals = pts.shape[1]
        hess = np.zeros(shape=(num_vals, num_vals))
        return hess