gosolnp: Random Initialization and Multiple Restarts of the solnp solver.

Description

When the objective function is non-smooth or has many local minima, it is hard to judge the optimality of the solution, and this usually depends critically on the starting parameters. This function enables the generation of a set of randomly chosen parameters from which to initialize multiple restarts of the solver (see note for details).

Usage

gosolnp(
  pars = NULL,
  fixed = NULL,
  fun,
  eqfun = NULL,
  eqB = NULL,
  ineqfun = NULL,
  ineqLB = NULL,
  ineqUB = NULL,
  LB = NULL,
  UB = NULL,
  control = list(),
  distr = rep(1, length(LB)),
  distr.opt = list(),
  n.restarts = 1,
  n.sim = 20000,
  cluster = NULL,
  rseed = NULL,
  ...
)

Value

A list containing the following values:

pars: Optimal Parameters.
convergence: Indicates whether the solver has converged (0) or not (1).
values: Vector of function values during optimization with last one the value at the optimal.
lagrange: The vector of Lagrange multipliers.
hessian: The Hessian at the optimal solution.
ineqx0: The estimated optimal inequality vector of slack variables used for transforming the inequality into an equality constraint.
nfuneval: The number of function evaluations.
elapsed: Time taken to compute solution.
start.pars: The parameter vector used to start the solver

Arguments

pars: The starting parameter vector. This is not required unless the fixed option is also used.
fixed: The numeric index which indicates those parameters which should stay fixed instead of being randomly generated.
fun: The main function which takes as first argument the parameter vector and returns a single value.
eqfun: (Optional) The equality constraint function returning the vector of evaluated equality constraints.
eqB: (Optional) The equality constraints.
ineqfun: (Optional) The inequality constraint function returning the vector of evaluated inequality constraints.
ineqLB: (Optional) The lower bound of the inequality constraints.
ineqUB: (Optional) The upper bound of the inequality constraints.
LB: The lower bound on the parameters. This is not optional in this function.
UB: The upper bound on the parameters. This is not optional in this function.
control: (Optional) The control list of optimization parameters. The eval.type option in this control list denotes whether to evaluate the function as is and exclude inequality violations in the final ranking (default, value = 1), else whether to evaluate a penalty barrier function comprised of the objective and all constraints (value = 2). See solnp function documentation for details of the remaining control options.
distr: A numeric vector of length equal to the number of parameters, indicating the choice of distribution to use for the random parameter generation. Choices are uniform (1), truncated normal (2), and normal (3).
distr.opt: If any choice in distr was anything other than uniform (1), this is a list equal to the length of the parameters with sub-components for the mean and sd, which are required in the truncated normal and normal distributions.
n.restarts: The number of solver restarts required.
n.sim: The number of random parameters to generate for every restart of the solver. Note that there will always be significant rejections if inequality bounds are present. Also, this choice should also be motivated by the width of the upper and lower bounds.
cluster: If you want to make use of parallel functionality, initialize and pass a cluster object from the parallel package (see details), and remember to terminate it!
rseed: (Optional) A seed to initiate the random number generator, else system time will be used.
...: (Optional) Additional parameters passed to the main, equality or inequality functions

Author

Alexios Galanos and Stefan Theussl
Y.Ye (original matlab version of solnp)

Details

Given a set of lower and upper bounds, the function generates, for those parameters not set as fixed, random values from one of the 3 chosen distributions. Depending on the eval.type option of the control argument, the function is either directly evaluated for those points not violating any inequality constraints, or indirectly via a penalty barrier function jointly comprising the objective and constraints. The resulting values are then sorted, and the best N (N = random.restart) parameter vectors (corresponding to the best N objective function values) chosen in order to initialize the solver. Since version 1.14, it is up to the user to prepare and pass a cluster object from the parallel package for use with gosolnp, after which the parLapply function is used. If your function makes use of additional packages, or functions, then make sure to export them via the clusterExport function of the parallel package. Additional arguments passed to the solver via the ... option are evaluated and exported by gosolnp to the cluster.

References

Y.Ye, Interior algorithms for linear, quadratic, and linearly constrained non linear programming, PhD Thesis, Department of EES Stanford University, Stanford CA.
Hu, X. and Shonkwiler, R. and Spruill, M.C. Random Restarts in Global Optimization, 1994, Georgia Institute of technology, Atlanta.

Examples

Run this code

if (FALSE) {
# [Example 1]
# Distributions of Electrons on a Sphere Problem:
# Given n electrons, find the equilibrium state distribution (of minimal Coulomb
# potential) of the electrons positioned on a conducting sphere. This model is
# from the COPS benchmarking suite. See http://www-unix.mcs.anl.gov/~more/cops/.
gofn = function(dat, n)
{

	x = dat[1:n]
	y = dat[(n+1):(2*n)]
	z = dat[(2*n+1):(3*n)]
	ii = matrix(1:n, ncol = n, nrow = n, byrow = TRUE)
	jj = matrix(1:n, ncol = n, nrow = n)
	ij = which(ii

Run the code above in your browser using DataLab