Control options for multistart() — multistartControl • nlmixr2extra

Control options for multistart()

Usage

multistartControl(
  n = 10L,
  nFit = NULL,
  sampling = c("uniform", "lhs", "normal"),
  spread = 0.2,
  which = NULL,
  perturbOmega = TRUE,
  omegaFold = 2,
  around = c("final", "initial"),
  screen = c("posthoc", "none"),
  refitBest = TRUE,
  excludeBoundary = TRUE,
  keepFits = TRUE,
  seed = 1234L,
  cores = 1L,
  cacheDir = NULL,
  restart = FALSE
)

Arguments

n: Number of candidate starting points to generate. The first candidate is always the unperturbed starting point, so n = 1 reproduces the original fit.
nFit: Number of candidates to fully estimate after screening. NULL (the default) estimates every candidate. Ignored when screen = "none".
sampling: How the starting points are drawn around the initial estimates: "uniform" (the default) draws uniformly within spread, "lhs" uses a Latin hypercube over the same interval so the range is covered more evenly, and "normal" draws normally with a standard deviation of spread.
spread: Fractional spread of the perturbation on the estimation scale. A parameter with initial estimate est is perturbed within est +/- spread*max(abs(est), 1).
which: Names of the population parameters to perturb; NULL (the default) perturbs every unfixed population parameter.
perturbOmega: Should the between-subject variability estimates be perturbed as well?
omegaFold: Fold-range for the omega perturbation. A variance v is drawn within v/omegaFold and v*omegaFold.
around: When starting from a fit, perturb around the fit's "final" estimates (the default, which asks "is this a local optimum?") or around the "initial" estimates the fit started from (which asks "how sensitive was this fit to where I started?").
screen: Cheap pre-selection of candidates. "posthoc" (the default) evaluates the objective function at each candidate with an empirical Bayes step only and fully estimates the best nFit; "none" estimates every candidate.
refitBest: Re-run the best start with the full estimation control, so the returned fit has the covariance step and tables the exploratory runs skip.
excludeBoundary: Should fits with a parameter at a boundary be excluded when picking the best start? Matches getMinAICFit().
keepFits: Keep each start's fit in the result. Setting this to FALSE keeps only the summary and the best fit, which is much smaller.
seed: Integer seed. The perturbations are drawn from this seed and each start is estimated with its own derived seed.
cores: Number of starts to estimate at once. Each estimation already uses every available thread internally, so the default of 1 is usually the fastest choice; see the "Parallel estimation" section of multistart().
cacheDir: Directory used to cache the individual starts so that an interrupted run can be resumed. NULL (the default) derives a name from the model; NA disables caching.
restart: Discard any cached results and start over.

Value

A validated list of control options for multistart()

Control options for `multistart()`

Usage

Arguments

Value

See also