Control Options for dynmodel

Source: R/dynmodel.R

Control Options for dynmodel

dynmodelControl(
  ...,
  ci = 0.95,
  nlmixrOutput = FALSE,
  digs = 3,
  lower = -Inf,
  upper = Inf,
  method = c("bobyqa", "Nelder-Mead", "lbfgsb3c", "L-BFGS-B", "PORT", "mma",
    "lbfgsbLG", "slsqp", "Rvmmin"),
  maxeval = 999,
  scaleTo = 1,
  scaleObjective = 0,
  normType = c("rescale2", "constant", "mean", "rescale", "std", "len"),
  scaleType = c("nlmixr", "norm", "mult", "multAdd"),
  scaleCmax = 1e+05,
  scaleCmin = 1e-05,
  scaleC = NULL,
  scaleC0 = 1e+05,
  atol = NULL,
  rtol = NULL,
  ssAtol = NULL,
  ssRtol = NULL,
  npt = NULL,
  rhobeg = 0.2,
  rhoend = NULL,
  iprint = 0,
  print = 1,
  maxfun = NULL,
  trace = 0,
  factr = NULL,
  pgtol = NULL,
  abstol = NULL,
  reltol = NULL,
  lmm = NULL,
  maxit = 100000L,
  eval.max = NULL,
  iter.max = NULL,
  abs.tol = NULL,
  rel.tol = NULL,
  x.tol = NULL,
  xf.tol = NULL,
  step.min = NULL,
  step.max = NULL,
  sing.tol = NULL,
  scale.init = NULL,
  diff.g = NULL,
  boundTol = NULL,
  epsilon = NULL,
  derivSwitchTol = NULL,
  sigdig = 4,
  covMethod = c("nlmixrHess", "optimHess"),
  gillK = 10L,
  gillStep = 4,
  gillFtol = 0,
  gillRtol = sqrt(.Machine$double.eps),
  gillKcov = 10L,
  gillStepCov = 2,
  gillFtolCov = 0,
  rxControl = NULL
)

Arguments

...

Other arguments including scaling factors for each compartment. This includes S# = numeric will scale a compartment # by a dividing the compartment amount by the scale factor, like NONMEM.

ci

Confidence level for some tables. By default this is 0.95 or 95% confidence.

nlmixrOutput

Option to change output style to nlmixr output. By default this is FALSE.

digs

Option for the number of significant digits of the output. By default this is 3.

lower

Lower bounds on the parameters used in optimization. By default this is -Inf.

upper

Upper bounds on the parameters used in optimization. By default this is Inf.

method

The method for solving ODEs. Currently this supports:

  • "liblsoda" thread safe lsoda. This supports parallel thread-based solving, and ignores user Jacobian specification.

  • "lsoda" -- LSODA solver. Does not support parallel thread-based solving, but allows user Jacobian specification.

  • "dop853" -- DOP853 solver. Does not support parallel thread-based solving nor user Jacobain specification

  • "indLin" -- Solving through inductive linearization. The RxODE dll must be setup specially to use this solving routine.

maxeval

Maximum number of iterations for Nelder-Mead of simplex search. By default this is 999.

scaleTo

Scale the initial parameter estimate to this value. By default this is 1. When zero or below, no scaling is performed.

scaleObjective

Scale the initial objective function to this value. By default this is 1.

normType

This is the type of parameter normalization/scaling used to get the scaled initial values for nlmixr. These are used with scaleType of.

With the exception of rescale2, these come from Feature Scaling. The rescale2 The rescaling is the same type described in the OptdesX software manual.

In general, all all scaling formula can be described by:

v_scaled = (v_unscaled-C_1)/C_2

Where

The other data normalization approaches follow the following formula

v_scaled = (v_unscaled-C_1)/C_2;

  • rescale2 This scales all parameters from (-1 to 1). The relative differences between the parameters are preserved with this approach and the constants are:

    C_1 = (max(all unscaled values)+min(all unscaled values))/2

    C_2 = (max(all unscaled values) - min(all unscaled values))/2

  • rescale or min-max normalization. This rescales all parameters from (0 to 1). As in the rescale2 the relative differences are preserved. In this approach:

    C_1 = min(all unscaled values)

    C_2 = max(all unscaled values) - min(all unscaled values)

  • mean or mean normalization. This rescales to center the parameters around the mean but the parameters are from 0 to 1. In this approach:

    C_1 = mean(all unscaled values)

    C_2 = max(all unscaled values) - min(all unscaled values)

  • std or standardization. This standardizes by the mean and standard deviation. In this approach:

    C_1 = mean(all unscaled values)

    C_2 = sd(all unscaled values)

  • len or unit length scaling. This scales the parameters to the unit length. For this approach we use the Euclidean length, that is:

    C_1 = 0

    C_2 = sqrt(v_1^2 + v_2^2 + ... + v_n^2)

  • constant which does not perform data normalization. That is

    C_1 = 0

    C_2 = 1

scaleType

The scaling scheme for nlmixr. The supported types are:

  • nlmixr In this approach the scaling is performed by the following equation:

    v_scaled = (v_current - v_init)/scaleC[i] + scaleTo

    The scaleTo parameter is specified by the normType, and the scales are specified by scaleC.

  • norm This approach uses the simple scaling provided by the normType argument.

  • mult This approach does not use the data normalization provided by normType, but rather uses multiplicative scaling to a constant provided by the scaleTo argument.

    In this case:

    v_scaled = v_current/v_init*scaleTo

  • multAdd This approach changes the scaling based on the parameter being specified. If a parameter is defined in an exponential block (ie exp(theta)), then it is scaled on a linearly, that is:

    v_scaled = (v_current-v_init) + scaleTo

    Otherwise the parameter is scaled multiplicatively.

    v_scaled = v_current/v_init*scaleTo

scaleCmax

Maximum value of the scaleC to prevent overflow.

scaleCmin

Minimum value of the scaleC to prevent underflow.

scaleC

The scaling constant used with scaleType=nlmixr. When not specified, it is based on the type of parameter that is estimated. The idea is to keep the derivatives similar on a log scale to have similar gradient sizes. Hence parameters like log(exp(theta)) would have a scaling factor of 1 and log(theta) would have a scaling factor of ini_value (to scale by 1/value; ie d/dt(log(ini_value)) = 1/ini_value or scaleC=ini_value)

  • For parameters in an exponential (ie exp(theta)) or parameters specifying powers, boxCox or yeoJohnson transformations , this is 1.

  • For additive, proportional, lognormal error structures, these are given by 0.5*abs(initial_estimate)

  • Factorials are scaled by abs(1/digamma(inital_estimate+1))

  • parameters in a log scale (ie log(theta)) are transformed by log(abs(initial_estimate))*abs(initial_estimate)

These parameter scaling coefficients are chose to try to keep similar slopes among parameters. That is they all follow the slopes approximately on a log-scale.

While these are chosen in a logical manner, they may not always apply. You can specify each parameters scaling factor by this parameter if you wish.

scaleC0

Number to adjust the scaling factor by if the initial gradient is zero.

atol

a numeric absolute tolerance (1e-8 by default) used by the ODE solver to determine if a good solution has been achieved; This is also used in the solved linear model to check if prior doses do not add anything to the solution.

rtol

a numeric relative tolerance (1e-6 by default) used by the ODE solver to determine if a good solution has been achieved. This is also used in the solved linear model to check if prior doses do not add anything to the solution.

ssAtol

Steady state atol convergence factor. Can be a vector based on each state.

ssRtol

Steady state rtol convergence factor. Can be a vector based on each state.

npt

The number of points used to approximate the objective function via a quadratic approximation for bobyqa. The value of npt must be in the interval [n+2,(n+1)(n+2)/2] where n is the number of parameters in par. Choices that exceed 2*n+1 are not recommended. If not defined, it will be set to 2*n + 1

rhobeg

Beginning change in parameters for bobyqa algorithm (trust region). By default this is 0.2 or 20 parameters when the parameters are scaled to 1. rhobeg and rhoend must be set to the initial and final values of a trust region radius, so both must be positive with 0 < rhoend < rhobeg. Typically rhobeg should be about one tenth of the greatest expected change to a variable. Note also that smallest difference abs(upper-lower) should be greater than or equal to rhobeg*2. If this is not the case then rhobeg will be adjusted.

rhoend

The smallest value of the trust region radius that is allowed. If not defined, then 10^(-sigdig-1) will be used.

iprint

Print option for optimization. See bobyqa, lbfgsb3c, and lbfgs for more details. By default this is 0.

print

Integer representing when the outer step is printed. When this is 0 or do not print the iterations. 1 is print every function evaluation (default), 5 is print every 5 evaluations.

maxfun

The maximum allowed number of function evaluations. If this is exceeded, the method will terminate. See bobyqa for more details. By default this value is NULL.

trace

Tracing information on the progress of the optimization is produced. See bobyqa, lbfgsb3c, and lbfgs for more details. By default this is 0.

factr

Controls the convergence of the "L-BFGS-B" method. Convergence occurs when the reduction in the objective is within this factor of the machine tolerance. Default is 1e10, which gives a tolerance of about 2e-6, approximately 4 sigdigs. You can check your exact tolerance by multiplying this value by .Machine$double.eps

pgtol

is a double precision variable.

On entry pgtol >= 0 is specified by the user. The iteration will stop when:

max(\| proj g_i \| i = 1, ..., n) <= lbfgsPgtol

where pg_i is the ith component of the projected gradient.

On exit pgtol is unchanged. This defaults to zero, when the check is suppressed.

abstol

Absolute tolerance for nlmixr optimizer

reltol

tolerance for nlmixr

lmm

An integer giving the number of BFGS updates retained in the "L-BFGS-B" method, It defaults to 7.

maxit

Maximum number of iterations for lbfgsb3c. See lbfgsb3c for more details. By default this is 100000L.

eval.max

Number of maximum evaluations of the objective function

iter.max

Maximum number of iterations allowed.

abs.tol

Used in Nelder-Mead optimization and PORT optimization. Absolute tolerance. Defaults to 0 so the absolute convergence test is not used. If the objective function is known to be non-negative, the previous default of 1e-20 would be more appropriate.

rel.tol

Relative tolerance before nlminb stops.

x.tol

X tolerance for nlmixr optimizers

xf.tol

Used in Nelder-Mead optimization and PORT optimization. false convergence tolerance. Defaults to 2.2e-14. See nlminb for more details.

step.min

Used in Nelder-Mead optimization and PORT optimization. Minimum step size. By default this is 1. See nlminb for more details.

step.max

Used in Nelder-Mead optimization and PORT optimization. Maximum step size. By default this is 1. See nlminb for more details.

sing.tol

Used in Nelder-Mead optimization and PORT optimization. Singular convergence tolerance; defaults to rel.tol. See nlminb for more details.

scale.init

Used in Nelder-Mead optimization and PORT optimization. See nlminb for more details.

diff.g

Used in Nelder-Mead optimization and PORT optimization. An estimated bound on the relative error in the objective function value. See nlminb for more details.

boundTol

Tolerance for boundary issues.

epsilon

Precision of estimate for n1qn1 optimization.

derivSwitchTol

The tolerance to switch forward to central differences.

sigdig

Optimization significant digits. This controls:

  • The tolerance of the inner and outer optimization is 10^-sigdig

  • The tolerance of the ODE solvers is 0.5*10^(-sigdig-2); For the sensitivity equations and steady-state solutions the default is 0.5*10^(-sigdig-1.5) (sensitivity changes only applicable for liblsoda)

  • The tolerance of the boundary check is 5 * 10 ^ (-sigdig + 1)

  • The significant figures that some tables are rounded to.

covMethod

Method for calculating covariance. In this discussion, R is the Hessian matrix of the objective function. The S matrix is the sum of individual gradient cross-product (evaluated at the individual empirical Bayes estimates).

gillK

The total number of possible steps to determine the optimal forward/central difference step size per parameter (by the Gill 1983 method). If 0, no optimal step size is determined. Otherwise this is the optimal step size determined.

gillStep

When looking for the optimal forward difference step size, this is This is the step size to increase the initial estimate by. So each iteration the new step size = (prior step size)*gillStep

gillFtol

The gillFtol is the gradient error tolerance that is acceptable before issuing a warning/error about the gradient estimates.

gillRtol

The relative tolerance used for Gill 1983 determination of optimal step size.

gillKcov

The total number of possible steps to determine the optimal forward/central difference step size per parameter (by the Gill 1983 method) during the covariance step. If 0, no optimal step size is determined. Otherwise this is the optimal step size determined.

gillStepCov

When looking for the optimal forward difference step size, this is This is the step size to increase the initial estimate by. So each iteration during the covariance step is equal to the new step size = (prior step size)*gillStepCov

gillFtolCov

The gillFtol is the gradient error tolerance that is acceptable before issuing a warning/error about the gradient estimates during the covariance step.

rxControl

This uses RxODE family of objects, file, or model specification to solve a ODE system. See rxControl for more details. By default this is NULL.

Value

dynmodelControl list for options during dynmodel optimization

Author

Mason McComb and Matthew L. Fidler