o a maximum rather than a minimum of maxiter starting values are 
  generated - todo

o if start is a data frame with n+1 rows where n is the number of
  parameters to optimize then Nelder Mead is used first - todo

o add minpack.lm interface - todo