RLS-OF 1.7

Regularised least squares objective function solver

This package provides trust region algorithms (TRA) for finding the minimum of some function. At the minute it contains only Levenberg-Marquart, but will be expanded to include NL2SOL and Powell's dogleg.

Levenberg-Marquardt

Example

An example is included within the package, simply call:

import TRA as TRA
def forward_model(x):
    y = np.array(x[0] ** 2 + x[1] ** 2)
    y = y.reshape((1, 1))
    return y

def compute_gradient(x):
    g = np.array(([2 * x[0]], [2 * x[1]]))
    g = g.reshape((2, 1))
    return g

def compute_hessian(x):
    h = np.array(([2, 0], [0, 2]))
    h = h.reshape((2, 2))
    return h


initial_prediction = np.array([5, 2.7])

LM_algorithm = TRA.Levenberg_Marquart(initial_prediction, compute_hessian, compute_gradient,
                                                  forward_model, d_param=1e-50,
                                                  lower_constraint=-np.inf,
                                                  upper_constraint=np.inf,
                                                  num_iterations=5)

minimum = LM_algorithm.optimisation_main()

This is a simple example, but shows how to use the Levenberg_Marquart class.

Function calls and arguments

There are a number of default values within the Levenberg_Marqaurdt class, including constraints on the solution, the number of iterations amd the damping parameter corresponding to the trust region. Three functions are required when instantiating a class object, one for computing the gradient, one for the Hessian and one for the mapping of the input to ouput (forward model).

:
def forward_model(x)
    :
    return f(x)
def compute_gradient(x):
    :
return grad

def compute_hessian(x):
    :
return hessian

initial_prediction = x0

LM_object = TRA.Levenberg_Marquart(initial_prediction, compute_hessian, compute_gradient,
                                                  forward_model, d_param=1e-50,
                                                  lower_constraint=-np.inf,
                                                  upper_constraint=np.inf,
                                                  num_iterations=5)

Theory

For the theory behind the code see [1] and [2].

References

[1] Jorge Nocedal and Stephen J. Wright (2006). Numerical Optimization.

[2] Andrew R. Conn, Nicholas I. M. Gould, and P.L. Toint (2000). Trust Region Methods.