nloptr

Introduction

NLopt addresses general nonlinear optimization problems of the form: \[ \begin{aligned} &\min_{x \in R^n} f(x) \\ s.t.& g(x) \leq 0 \\ & h(x) = 0 \\ & x_L \leq x \leq x_U \end{aligned} \]

where \(f(\cdot)\) is the objective function and \(x\) represents the \(n\) optimization parameters. This problem may optionally be subject to the bound constraints (also called box constraints), \(x_L\) and \(x_U\). For partially or totally unconstrained problems the bounds can take values \(-\infty\) or \(\infty\). One may also optionally have \(m\) nonlinear inequality constraints—sometimes called a nonlinear programming problem—which may be specified in \(g(\cdot)\), and equality constraints which may be specified in \(h(\cdot)\). Note that not all of the algorithms in NLopt can handle constraints.

This vignette describes how to formulate minimization problems to be solved with the R interface to NLopt. If you want to use the C interface directly or are interested in the Matlab interface, there are other sources of documentation available. Some of the information here has been taken from the NLopt website, where more details are available. All credit for implementing the C code for the different algorithms available in NLopt should go to the respective authors. Also, please see the website for information on how to cite NLopt and the algorithms you use.

Installation

This package is on CRAN and can be installed from within R using

install.packages("nloptr")

or

install.packages("nloptr", type = "source")

to install the package from source. You should now be able to load the R interface to NLopt and read the help.

library("nloptr")
?nloptr

The most recent experimental source version of nloptr can be installed from Github using the remotes package:

# install.packages("remotes")
remotes::install_github("astamm/nloptr")

Minimizing the Rosenbrock Banana function

As a first example we will solve an unconstrained minimization problem. The function we look at is the Rosenbrock Banana function: \[ f(x) = 100 \left(x_2-x_1^2\right)^2 + \left(1-x_1\right)^2, \]

which is also used as an example in the documentation for the standard R optimizer optim. The gradient of the objective function is given by: \[ \nabla f(x) = \left(\begin{array}[1]{c} -400 \cdot x_1 \cdot (x_2 - x_1^2) - 2 \cdot (1 - x_1) \\ 200 \cdot (x_2 - x_1^2) \end{array} \right). \]

Not all of the algorithms in NLopt need gradients to be supplied by the user. We will show examples with and without supplying the gradient. After loading the library.

library(nloptr)

We start by specifying the objective function and its gradient:

## Rosenbrock Banana function
eval_f <- function(x) {
  100 * (x[2] - x[1] * x[1]) ^ 2 + (1 - x[1]) ^ 2
}

## Gradient of Rosenbrock Banana function
eval_grad_f <- function(x) {
  c(-400 * x[1] * (x[2] - x[1] * x[1]) - 2 * (1 - x[1]),
    200 * (x[2] - x[1] * x[1]))
}

We define initial values

# initial values
x0 <- c(-1.2, 1)

and then minimize the function using the nloptr command. This command runs some checks on the supplied inputs and returns an object with the exit code of the solver, the optimal value of the objective function and the solution. Before we can minimize the function we need to specify which algorithm we want to use

opts <- list("algorithm" = "NLOPT_LD_LBFGS",
             "xtol_rel" = 1.0e-8)

Here we use the L-BFGS algorithm (Nocedal 1980; Liu and Nocedal 1989). The characters LD in the algorithm show that this algorithm looks for local minima (L) using a derivative-based (D) algorithm. Other algorithms look for global (G) minima, or they don’t need derivatives (N). We also specified the termination criterium in terms of the relative x-tolerance. Other termination criteria are available (see Appendix \ref{sec:descoptions} for a full list of options). We then solve the minimization problem using

# solve Rosenbrock Banana function
res <- nloptr(x0 = x0,
              eval_f = eval_f,
              eval_grad_f = eval_grad_f,
              opts = opts)

We can see the results by printing the resulting object.

print(res)

## 
## Call:
## 
## nloptr(x0 = x0, eval_f = eval_f, eval_grad_f = eval_grad_f, opts = opts)
## 
## 
## 
## Minimization using NLopt version 2.7.1 
## 
## NLopt solver status: 1 ( NLOPT_SUCCESS: Generic success 
## return value. )
## 
## Number of Iterations....: 56 
## Termination conditions:  xtol_rel: 1e-08 
## Number of inequality constraints:  0 
## Number of equality constraints:    0 
## Optimal value of objective function:  5.6213401034694e-23 
## Optimal value of controls: 1 1

Sometimes the objective function and its gradient contain common terms. To economize on calculations, we can return the objective and its gradient in a list. For the Rosenbrock Banana function we have for instance:

## Rosenbrock Banana function and gradient in one function
eval_f_list <- function(x) {
  common_term <- x[2] - x[1] * x[1]
  list("objective" = 100 * common_term ^ 2 + (1 - x[1]) ^ 2,
       "gradient"  = c(-400 * x[1] * common_term - 2 * (1 - x[1]),
                       200 * common_term))
}

which we minimize using

res <- nloptr(x0 = x0,
              eval_f = eval_f_list,
              opts = opts)
print(res)

## 
## Call:
## nloptr(x0 = x0, eval_f = eval_f_list, opts = opts)
## 
## 
## Minimization using NLopt version 2.7.1 
## 
## NLopt solver status: 1 ( NLOPT_SUCCESS: Generic success 
## return value. )
## 
## Number of Iterations....: 56 
## Termination conditions:  xtol_rel: 1e-08 
## Number of inequality constraints:  0 
## Number of equality constraints:    0 
## Optimal value of objective function:  5.6213401034694e-23 
## Optimal value of controls: 1 1

This gives the same results as before.

Minimization with inequality constraints

This section shows how to minimize a function subject to inequality constraints. This example is the same as the one used in the tutorial on the NLopt website. The problem we want to solve is:

\[ \begin{aligned} &\min_{x \in R^n} \sqrt{x_2} \\ s.t.& x_2 \geq 0 \\ & x_2 \geq (a_1 x_1 + b_1)^3 \\ & x_2 \geq (a_2 x_1 + b_2)^3, \end{aligned} \]

where \(a_1 = 2\), \(b_1 = 0\), \(a_2 = -1\), and \(b_2 = 1\). In order to solve this problem, we first have to re-formulate the constraints to be of the form \(g(x) \leq 0\). Note that the first constraint is a bound on \(x_2\), which we will add later. The other two constraints can be re-written as:

\[ \begin{aligned} (a_1 x_1 + b_1)^3 - x_2 &\leq 0 \\ (a_2 x_1 + b_2)^3 - x_2 &\leq 0 \end{aligned} \]

First, define R functions to calculate the objective function and its gradient:

# objective function
eval_f0 <- function(x, a, b) {
  sqrt(x[2])
}

# gradient of objective function
eval_grad_f0 <- function(x, a, b) {
  c(0, 0.5 / sqrt(x[2]))
}

If needed, these can of course be calculated in the same function as before. Then define the two constraints and the Jacobian of the constraints:

# constraint function
eval_g0 <- function(x, a, b) {
  (a * x[1] + b) ^ 3 - x[2]
}

# Jacobian of constraint
eval_jac_g0 <- function(x, a, b) {
  rbind(c(3 * a[1] * (a[1] * x[1] + b[1]) ^ 2, -1.0),
        c(3 * a[2] * (a[2] * x[1] + b[2]) ^ 2, -1.0))
}

Note that all of the functions above depend on additional parameters, a and b. We have to supply specific values for these when we invoke the optimization command. The constraint function eval_g0 returns a vector with in this case the same length as the vectors a and b. The function calculating the Jacobian of the constraint should return a matrix where the number of rows equal the number of constraints (in this case two). The number of columns should equal the number of control variables (two in this case as well).

After defining values for the parameters

# define parameters
a <- c(2, -1)
b <- c(0, 1)

we can minimize the function subject to the constraints with the following command:

# Solve using NLOPT_LD_MMA with gradient information supplied in separate
# function
res0 <- nloptr(x0 = c(1.234, 5.678),
               eval_f = eval_f0,
               eval_grad_f = eval_grad_f0,
               lb = c(-Inf, 0),
               ub = c(Inf, Inf),
               eval_g_ineq = eval_g0,
               eval_jac_g_ineq = eval_jac_g0,
               opts = list("algorithm" = "NLOPT_LD_MMA",
                           "xtol_rel" = 1.0e-8,
                           "print_level" = 2,
                           "check_derivatives" = TRUE,
                           "check_derivatives_print" = "all"),
               a = a,
               b = b)

## Checking gradients of objective function.

## Derivative checker results: 0 error(s) detected.

## 
##   eval_grad_f[1] = 0.000000e+00 ~ 0.000000e+00   [0.000000e+00]
##   eval_grad_f[2] = 2.098323e-01 ~ 2.098323e-01   [1.422937e-09]

## Checking gradients of inequality constraints.

## Derivative checker results: 0 error(s) detected.

## 
##   eval_jac_g_ineq[1, 1] =  3.654614e+01 ~  3.654614e+01   [1.667794e-08]
##   eval_jac_g_ineq[2, 1] = -1.642680e-01 ~ -1.642680e-01   [2.103453e-07]
##   eval_jac_g_ineq[1, 2] = -1.000000e+00 ~ -1.000000e+00   [0.000000e+00]
##   eval_jac_g_ineq[2, 2] = -1.000000e+00 ~ -1.000000e+00   [0.000000e+00]

## iteration: 1
##  f(x) = 2.382855
##  g(x) = (9.354647, -5.690813)
## iteration: 2
##  f(x) = 2.356135
##  g(x) = (-0.122988, -5.549587)
## iteration: 3
##  f(x) = 2.245864
##  g(x) = (-0.531886, -5.038655)
## iteration: 4
##  f(x) = 2.019102
##  g(x) = (-3.225103, -3.931195)
## iteration: 5
##  f(x) = 1.740934
##  g(x) = (-2.676263, -2.761136)
## iteration: 6
##  f(x) = 1.404206
##  g(x) = (-1.674056, -1.676216)
## iteration: 7
##  f(x) = 1.022295
##  g(x) = (-0.748790, -0.748792)
## iteration: 8
##  f(x) = 0.685203
##  g(x) = (-0.173206, -0.173207)
## iteration: 9
##  f(x) = 0.552985
##  g(x) = (-0.009496, -0.009496)
## iteration: 10
##  f(x) = 0.544354
##  g(x) = (-0.000025, -0.000025)
## iteration: 11
##  f(x) = 0.544331
##  g(x) = (0.000000, 0.000000)
## iteration: 12
##  f(x) = 0.544331
##  g(x) = (0.000000, 0.000000)
## iteration: 13
##  f(x) = 0.544331
##  g(x) = (0.000000, 0.000000)
## iteration: 14
##  f(x) = 0.544331
##  g(x) = (0.000000, 0.000000)
## iteration: 15
##  f(x) = 0.544331
##  g(x) = (0.000000, 0.000000)
## iteration: 16
##  f(x) = 0.544331
##  g(x) = (0.000000, 0.000000)
## iteration: 17
##  f(x) = 0.544331
##  g(x) = (0.000000, 0.000000)
## iteration: 18
##  f(x) = 0.544331
##  g(x) = (0.000000, 0.000000)

print(res0)

## 
## Call:
## 
## nloptr(x0 = c(1.234, 5.678), eval_f = eval_f0, eval_grad_f = eval_grad_f0, 
##     lb = c(-Inf, 0), ub = c(Inf, Inf), eval_g_ineq = eval_g0, 
##     eval_jac_g_ineq = eval_jac_g0, opts = list(algorithm = "NLOPT_LD_MMA", 
##         xtol_rel = 1e-08, print_level = 2, check_derivatives = TRUE, 
##         check_derivatives_print = "all"), a = a, b = b)
## 
## 
## Minimization using NLopt version 2.7.1 
## 
## NLopt solver status: 4 ( NLOPT_XTOL_REACHED: Optimization 
## stopped because xtol_rel or xtol_abs (above) was reached. 
## )
## 
## Number of Iterations....: 18 
## Termination conditions:  xtol_rel: 1e-08 
## Number of inequality constraints:  2 
## Number of equality constraints:    0 
## Optimal value of objective function:  0.544331047591509 
## Optimal value of controls: 0.3333333 0.2962963

Here we supplied lower bounds for \(x_2\) in lb. There are no upper bounds for both control variables, so we supply Inf values. If we don’t supply lower or upper bounds, plus or minus infinity is chosen by default. The inequality constraints and its Jacobian are defined using eval_g_ineq and eval_jac_g_ineq. Not all algorithms can handle inequality constraints, so we have to specify one that does, NLOPT_LD_MMA (Svanberg 2002).

We also specify the option print_level to obtain output during the optimization process. For the available print_level values, see ?nloptr. Setting the check_derivatives option to TRUE, compares the gradients supplied by the user with a finite difference approximation in the initial point (x0). When this check is run, the option check_derivatives_print can be used to print all values of the derivative checker (all (default)), only those values that result in an error (errors) or no output (none), in which case only the number of errors is shown. The tolerance that determines if a difference between the analytic gradient and the finite difference approximation results in an error can be set using the option check_derivatives_tol (default = 1e-04). The first column shows the value of the analytic gradient, the second column shows the value of the finite difference approximation, and the third column shows the relative error. Stars are added at the front of a line if the relative error is larger than the specified tolerance.

Finally, we add all the parameters that have to be passed on to the objective and constraint functions, a and b.

We can also use a different algorithm to solve the same minimization problem. The only thing we have to change is the algorithm that we want to use, in this case NLOPT_LN_COBYLA, which is an algorithm that doesn’t need gradient information (Powell 1994, 1998).

# Solve using NLOPT_LN_COBYLA without gradient information
res1 <- nloptr(x0 = c(1.234, 5.678),
               eval_f = eval_f0,
               lb = c(-Inf, 0),
               ub = c(Inf, Inf),
               eval_g_ineq = eval_g0,
               opts = list("algorithm" = "NLOPT_LN_COBYLA",
                           "xtol_rel" = 1.0e-8),
               a = a,
               b = b)
print(res1)

## 
## Call:
## 
## nloptr(x0 = c(1.234, 5.678), eval_f = eval_f0, lb = c(-Inf, 0), 
##     ub = c(Inf, Inf), eval_g_ineq = eval_g0, opts = list(algorithm = "NLOPT_LN_COBYLA", 
##         xtol_rel = 1e-08), a = a, b = b)
## 
## 
## Minimization using NLopt version 2.7.1 
## 
## NLopt solver status: 4 ( NLOPT_XTOL_REACHED: Optimization 
## stopped because xtol_rel or xtol_abs (above) was reached. 
## )
## 
## Number of Iterations....: 50 
## Termination conditions:  xtol_rel: 1e-08 
## Number of inequality constraints:  2 
## Number of equality constraints:    0 
## Optimal value of objective function:  0.544331053951819 
## Optimal value of controls: 0.3333333 0.2962963

Derivative checker

The derivative checker can be called when supplying a minimization problem to nloptr, using the options check_derivatives, check_derivatives_tol and check_derivatives_print, but it can also be used separately. For example, define the function g, with vector outcome, and its gradient g_grad:

g <- function(x, a) {
  c(x[1] - a[1],
    x[2] - a[2],
    (x[1] - a[1]) ^ 2,
    (x[2] - a[2]) ^ 2,
    (x[1] - a[1]) ^ 3,
    (x[2] - a[2]) ^ 3)
}

g_grad <- function(x, a) {
  rbind(
    c(1, 0),
    c(0, 1),
    c(2 * (x[1] - a[1]), 0),
    c(2 * (x[1] - a[1]), 2 * (x[2] - a[2])),
    c(3 * (x[1] - a[2]) ^ 2, 0),
    c(0, 3 * (x[2] - a[2]) ^ 2)
  )
}

a is some vector containing data. The gradient contains some errors in this case. By calling the function check.derivatives we can check the user-supplied analytic gradients with a finite difference approximation at a point .x.

res <- check.derivatives(
  .x = c(1, 2),
  func = g,
  func_grad = g_grad,
  check_derivatives_print = "all",
  a = c(0.3, 0.8)
)

## Derivative checker results: 2 error(s) detected.

## 
##   grad_f[1, 1] = 1.00e+00 ~ 1.00e+00   [0.000000e+00]
##   grad_f[2, 1] = 0.00e+00 ~ 0.00e+00   [0.000000e+00]
##   grad_f[3, 1] = 1.40e+00 ~ 1.40e+00   [9.579318e-09]
## * grad_f[4, 1] = 1.40e+00 ~ 0.00e+00   [1.400000e+00]
## * grad_f[5, 1] = 1.20e-01 ~ 1.47e+00   [9.183673e-01]
##   grad_f[6, 1] = 0.00e+00 ~ 0.00e+00   [0.000000e+00]
##   grad_f[1, 2] = 0.00e+00 ~ 0.00e+00   [0.000000e+00]
##   grad_f[2, 2] = 1.00e+00 ~ 1.00e+00   [0.000000e+00]
##   grad_f[3, 2] = 0.00e+00 ~ 0.00e+00   [0.000000e+00]
##   grad_f[4, 2] = 2.40e+00 ~ 2.40e+00   [1.179675e-08]
##   grad_f[5, 2] = 0.00e+00 ~ 0.00e+00   [0.000000e+00]
##   grad_f[6, 2] = 4.32e+00 ~ 4.32e+00   [2.593906e-08]

The errors are shown on screen, where the option check_derivatives_print determines the amount of output you see. The value of the analytic gradient and the value of the finite difference approximation at the supplied point is returned in a list.

res

## $analytic
##      [,1] [,2]
## [1,] 1.00 0.00
## [2,] 0.00 1.00
## [3,] 1.40 0.00
## [4,] 1.40 2.40
## [5,] 0.12 0.00
## [6,] 0.00 4.32
## 
## $finite_difference
##      [,1] [,2]
## [1,] 1.00 0.00
## [2,] 0.00 1.00
## [3,] 1.40 0.00
## [4,] 0.00 2.40
## [5,] 1.47 0.00
## [6,] 0.00 4.32
## 
## $relative_error
##              [,1]         [,2]
## [1,] 0.000000e+00 0.000000e+00
## [2,] 0.000000e+00 0.000000e+00
## [3,] 9.579318e-09 0.000000e+00
## [4,] 1.400000e+00 1.179675e-08
## [5,] 9.183673e-01 0.000000e+00
## [6,] 0.000000e+00 2.593906e-08
## 
## $flag_derivative_warning
##       [,1]  [,2]
## [1,] FALSE FALSE
## [2,] FALSE FALSE
## [3,] FALSE FALSE
## [4,]  TRUE FALSE
## [5,]  TRUE FALSE
## [6,] FALSE FALSE

Note that not all errors will be picked up by the derivative checker. For instance, if we run the check with a = c(.5, .5), one of the errors is not flagged as an error.

Notes

The .R scripts in the tests directory contain more examples. For instance, hs071.R and systemofeq.R show how to solve problems with equality constraints. See the NLopt website for more details. Please let us know if there are any of features implemented in NLopt which should be implemented in nloptr.

Sometimes the optimization procedure terminates with a message maxtime was reached without evaluating the objective function. Submitting the same problem again usually solves this problem.

Description of options

nloptr::nloptr.print.options()

## algorithm
##  possible values: NLOPT_GN_DIRECT, NLOPT_GN_DIRECT_L,
##           NLOPT_GN_DIRECT_L_RAND, NLOPT_GN_DIRECT_NOSCAL,
##           NLOPT_GN_DIRECT_L_NOSCAL,
##           NLOPT_GN_DIRECT_L_RAND_NOSCAL,
##           NLOPT_GN_ORIG_DIRECT, NLOPT_GN_ORIG_DIRECT_L,
##           NLOPT_GD_STOGO, NLOPT_GD_STOGO_RAND,
##           NLOPT_LD_SLSQP, NLOPT_LD_LBFGS_NOCEDAL,
##           NLOPT_LD_LBFGS, NLOPT_LN_PRAXIS, NLOPT_LD_VAR1,
##           NLOPT_LD_VAR2, NLOPT_LD_TNEWTON,
##           NLOPT_LD_TNEWTON_RESTART,
##           NLOPT_LD_TNEWTON_PRECOND,
##           NLOPT_LD_TNEWTON_PRECOND_RESTART,
##           NLOPT_GN_CRS2_LM, NLOPT_GN_MLSL, NLOPT_GD_MLSL,
##           NLOPT_GN_MLSL_LDS, NLOPT_GD_MLSL_LDS,
##           NLOPT_LD_MMA, NLOPT_LD_CCSAQ, NLOPT_LN_COBYLA,
##           NLOPT_LN_NEWUOA, NLOPT_LN_NEWUOA_BOUND,
##           NLOPT_LN_NELDERMEAD, NLOPT_LN_SBPLX,
##           NLOPT_LN_AUGLAG, NLOPT_LD_AUGLAG,
##           NLOPT_LN_AUGLAG_EQ, NLOPT_LD_AUGLAG_EQ,
##           NLOPT_LN_BOBYQA, NLOPT_GN_ISRES
##  default value:   none
## 
##  This option is required. Check the NLopt website for a description of
##  the algorithms.
## 
## stopval
##  possible values: -Inf <= stopval <= Inf
##  default value:   -Inf
## 
##  Stop minimization when an objective value <= stopval is found.
##  Setting stopval to -Inf disables this stopping criterion (default).
## 
## ftol_rel
##  possible values: ftol_rel > 0
##  default value:   0.0
## 
##  Stop when an optimization step (or an estimate of the optimum)
##  changes the objective function value by less than ftol_rel multiplied
##  by the absolute value of the function value. If there is any chance
##  that your optimum function value is close to zero, you might want to
##  set an absolute tolerance with ftol_abs as well. Criterion is
##  disabled if ftol_rel is non-positive (default).
## 
## ftol_abs
##  possible values: ftol_abs > 0
##  default value:   0.0
## 
##  Stop when an optimization step (or an estimate of the optimum)
##  changes the function value by less than ftol_abs. Criterion is
##  disabled if ftol_abs is non-positive (default).
## 
## xtol_rel
##  possible values: xtol_rel > 0
##  default value:   1.0e-04
## 
##  Stop when an optimization step (or an estimate of the optimum)
##  changes every parameter by less than xtol_rel multiplied by the
##  absolute value of the parameter. If there is any chance that an
##  optimal parameter is close to zero, you might want to set an absolute
##  tolerance with xtol_abs as well. Criterion is disabled if xtol_rel is
##  non-positive.
## 
## xtol_abs
##  possible values: xtol_abs > 0
##  default value:   rep(0.0, length(x0))
## 
##  xtol_abs is a vector of length n (the number of elements in x) giving
##  the tolerances: stop when an optimization step (or an estimate of the
##  optimum) changes every parameter x[i] by less than xtol_abs[i].
##  Criterion is disabled if all elements of xtol_abs are non-positive
##  (default).
## 
## maxeval
##  possible values: maxeval is a positive integer
##  default value:   100
## 
##  Stop when the number of function evaluations exceeds maxeval. This is
##  not a strict maximum: the number of function evaluations may exceed
##  maxeval slightly, depending upon the algorithm. Criterion is disabled
##  if maxeval is non-positive.
## 
## maxtime
##  possible values: maxtime > 0
##  default value:   -1.0
## 
##  Stop when the optimization time (in seconds) exceeds maxtime. This is
##  not a strict maximum: the time may exceed maxtime slightly, depending
##  upon the algorithm and on how slow your function evaluation is.
##  Criterion is disabled if maxtime is non-positive (default).
## 
## tol_constraints_ineq
##  possible values: tol_constraints_ineq > 0.0
##  default value:   rep(1e-8, num_constraints_ineq)
## 
##  The parameter tol_constraints_ineq is a vector of tolerances. Each
##  tolerance corresponds to one of the inequality constraints. The
##  tolerance is used for the purpose of stopping criteria only: a point
##  x is considered feasible for judging whether to stop the optimization
##  if eval_g_ineq(x) <= tol. A tolerance of zero means that NLopt will
##  try not to consider any x to be converged unless eval_g_ineq(x) is
##  strictly non-positive; generally, at least a small positive tolerance
##  is advisable to reduce sensitivity to rounding errors. By default the
##  tolerances for all inequality constraints are set to 1e-8.
## 
## tol_constraints_eq
##  possible values: tol_constraints_eq > 0.0
##  default value:   rep(1e-8, num_constraints_eq)
## 
##  The parameter tol_constraints_eq is a vector of tolerances. Each
##  tolerance corresponds to one of the equality constraints. The
##  tolerance is used for the purpose of stopping criteria only: a point
##  x is considered feasible for judging whether to stop the optimization
##  if abs(eval_g_ineq(x)) <= tol. For equality constraints, a small
##  positive tolerance is strongly advised in order to allow NLopt to
##  converge even if the equality constraint is slightly nonzero. By
##  default the tolerances for all quality constraints are set to 1e-8.
## 
## print_level
##  possible values: 0, 1, 2, or 3
##  default value:   0
## 
##  The option print_level controls how much output is shown during the
##  optimization process. Possible values: 0 (default): no output; 1:
##  show iteration number and value of objective function; 2: 1 + show
##  value of (in)equalities; 3: 2 + show value of controls.
## 
## check_derivatives
##  possible values: TRUE or FALSE
##  default value:   FALSE
## 
##  The option check_derivatives can be activated to compare the
##  user-supplied analytic gradients with finite difference
##  approximations.
## 
## check_derivatives_tol
##  possible values: check_derivatives_tol > 0.0
##  default value:   1e-04
## 
##  The option check_derivatives_tol determines when a difference between
##  an analytic gradient and its finite difference approximation is
##  flagged as an error.
## 
## check_derivatives_print
##  possible values: 'none', 'all', 'errors',
##  default value:   all
## 
##  The option check_derivatives_print controls the output of the
##  derivative checker (if check_derivatives == TRUE). All comparisons
##  are shown ('all'), only those comparisions that resulted in an error
##  ('error'), or only the number of errors is shown ('none').
## 
## print_options_doc
##  possible values: TRUE or FALSE
##  default value:   FALSE
## 
##  If TRUE, a description of all options and their current and default
##  values is printed to the screen.
## 
## population
##  possible values: population is a positive integer
##  default value:   0
## 
##  Several of the stochastic search algorithms (e.g., CRS, MLSL, and
##  ISRES) start by generating some initial population of random points
##  x. By default, this initial population size is chosen heuristically
##  in some algorithm-specific way, but the initial population can by
##  changed by setting a positive integer value for population. A
##  population of zero implies that the heuristic default will be used.
## 
## vector_storage
##  possible values: vector_storage is a positive integer
##  default value:   20
## 
##  Number of gradients to remember from previous optimization steps.
## 
## ranseed
##  possible values: ranseed is a positive integer
##  default value:   0
## 
##  For stochastic optimization algorithms, pseudorandom numbers are
##  generated. Set the random seed using ranseed if you want to use a
##  'deterministic' sequence of pseudorandom numbers, i.e. the same
##  sequence from run to run. If ranseed is 0 (default), the seed for the
##  random numbers is generated from the system time, so that you will
##  get a different sequence of pseudorandom numbers each time you run
##  your program.