Skip to contents

Convergence control

Source: R/rate.R
tda_control.Rd

Convergence control

Usage

tda_control(
  algorithm = NULL,
  tolsg = NULL,
  tolsp = NULL,
  tolg = NULL,
  tolf = NULL,
  maxit = NULL,
  tol_param = NULL,
  tol_var = NULL,
  tol_reduction = NULL,
  step_length = NULL,
  step_reduction = NULL,
  step_min = NULL,
  maxit_line = NULL,
  criterion = NULL,
  derivatives = NULL
)

tda_strict(tol = 1e-12)

tda_converged(x)

Arguments

algorithm

TDA minimiser number, its mina=: 1 direct search, 2 simplex, 3 conjugate gradients, 4 BFGS, 5 Newton (I, TDA's own default), 6 Newton (II), 7 CES with a quadratic model, 8 CES with a tensor model. 7 and 8 use the scale-invariant criterion.

tolsg, tolsp

tolerances for the scaled gradient and scaled parameter change, used by algorithms 7 and 8.

tolg, tolf

tolerances for the gradient norm and the function value, used by the default algorithm.

maxit

iteration cap, TDA's mxit=.

tol_param, tol_var, tol_reduction

three more tolerances TDA's minimiser reads for other algorithms/situations – tolp= (parameter change, direct search and others), tolv= (Simplex), tols= (reduction factor, direct search).

step_length, step_reduction, step_min

step-size controls for TDA's line search – slen=, sred=, smin=.

maxit_line

cap on line-search iterations within one step – TDA's mxitl=, default 50.

criterion

which convergence test to apply – TDA's crit=, 1 to 3. Confirmed limited: it cannot select the scale-invariant Dennis-Schnabel test described under Criteria above, which is reachable only through algorithm = 7 or 8 directly, not through this argument.

derivatives

0 for analytical derivatives (TDA's default when unset), 1 or 2 for numerical approximations of varying cost – TDA's dopt=. Applies only to algorithm = 7 or 8; harmless but inert for every other algorithm.

tol

tolerance for tda_strict, applied to both tolsg and tolsp.

x

a fitted model.

Value

A list of options for control=; tda_converged returns TRUE, FALSE, or NA where TDA reports nothing.

Criteria

TDA's default minimiser stops on the norm of the gradient against an absolute tolerance, plus a loose disjunction that also stops when the change in function value is small. Algorithms 7 and 8 switch to optstp(), the Dennis-Schnabel test on the largest relative gradient and relative parameter step. That is scale-invariant, closer to what R's optimisers do, and gives results that reproduce across platforms. It is reachable only through algorithm: TDA's crit= accepts 1 to 3 and cannot select it.

The default is TDA's, so a plain call reproduces a plain command file. tda_strict() opts in.

A caution

Algorithms 7 and 8 also lower the iteration cap to 20 when derivatives are numerical, against 100 otherwise, so asking for a tighter tolerance can turn a converged fit into a truncated one. Every fit reports how it stopped, warns if it did not converge, and tda_converged() reads that back.

See also

Other rate models: tda_constrain(), tda_dple(), tda_rates(), tda_split(), tda_survivor(), tda_transitions(), vcov.tda_fit()

Examples

set.seed(36)
d <- data.frame(x = rnorm(200))
d$y <- rbinom(200, 1, plogis(0.3 + 0.8 * d$x))
f <- tda_qreg(y ~ x, d, control = tda_control(maxit = 50, tolg = 1e-8))
coef(f)
#> Intercept         x 
#> 0.3851079 0.9164042 
tda_converged(f)
#> [1] TRUE

# tda_strict: the same idea, one shared tolerance for both algorithm 7/8
# criteria at once
f2 <- tda_qreg(y ~ x, d, control = tda_strict())
coef(f2)
#> Intercept         x 
#> 0.3851079 0.9164042