ms.control(maxiter = 50, tolerance, maxfcalls = 200,
          rel.tolerance, f.tolerance, failure = 1,
          scale, minscale, trace = F, flags, opt.parameters)

maxiter:

maximum number of iterations in the fitting algorithm. Default is 50.

tolerance:

if the change in the scaled parameter estimates is less than this value, the iteration is judged to have converged. The parameters are scaled by the algorithm in the internal calculations. Roughly, you can think of this tolerance as a relative convergence criterion in the parameters that should be approximately invariant to rescaling. Default is sqrt(.Machine$double.eps), where eps is the machine precision.

maxfcalls:

maximum number of times the expression for the function should be evaluated. Default is 200.

rel.tolerance:

if the relative change in the objective function is less than this value, the iteration is judged to have converged. Default max(10^-10, .Machine$double.eps^(2/3)).

f.tolerance:

if the absolute value of the objective function is less than this value, the iteration is judged to have converged. (Only relevant if the computed value at the minimum could be a hard zero.) Default 10^(-20).

failure:

what action should occur if the iteration fails? The default is a warning. If failure is 0, no warning is issued. If failure is 2 or more, a fatal error occurs.

scale:

vector of scale values, which is multiplied into the parameter vector for purposes of testing convergence and for other internal calculations. The algorithm does its own attempt at choosing a scale vector if none is provided.

minscale:

the optimization algorithm tries at each iteration to make the function decrease by stepping out along a suggested direction. If the function does not decrease, the scale of the step is cut back. If the scale drops below the minscale value (relative to an initial value of 1), the iteration is judged to have failed. Default is 100 * .Machine$double.eps.

trace:

logical value: if TRUE special trace printing for the iterative minimization is produced.

flags:

the integer and double-precision arrays of internal parameters to the minimization algorithm. These are returned in the fitted object (see ms.object). The strong of heart can consult the documentation for the underlying FORTRAN algorithm, in the PORT library, and change individual values. Note that these block settings will be overridden by any specific use of the other arguments above.

opt.parameters:

the integer and double-precision arrays of internal parameters to the minimization algorithm. These are returned in the fitted object (see ms.object). The strong of heart can consult the documentation for the underlying FORTRAN algorithm, in the PORT library, and change individual values. Note that these block settings will be overridden by any specific use of the other arguments above.

Chambers, J. M., and Hastie, T. J. (eds) (1990). Statistical Models in S, Chapter 10, "Nonlinear Models", Pacific Grove, CA.

lprob <- function(lp)log(1+ exp(lp)) - lp  # log-likelihood
fit.alpha <- ms(~ lprob(D * alpha), pingpong,
                control = list(maxiter = 10, tol = .001))