Specifying Fixed Parameters


In finding optimal parameters in nonlinear optimization and nonlinear least squares problems, we frequently wish to fix one or more parameters while allowing the rest to be adjusted to explore or optimize an objective function.

This vignette discusses some ideas about specifying the fixed parameters. A lot of the material is drawn from Nash J C (2014) Nonlinear parameter optimization using R tools Chichester UK: Wiley, in particular chapters 11 and 12.


Here are some of the ways fixed parameters may be specified in R packages.

From nlxb in package nlsr or previously in nlxb in package nlmrt.


Character vector of quoted parameter names. These parameters will NOT be altered by the algorithm. This approach has a simplicity that is attractive, but introduces an extra argument to calling sequences.

From nlfb


?? This needs to be changed to use names!?? Vector if indices of the parameters to be masked. These parameters will NOT be altered by the algorithm. Note that the mechanism here is different from that in nlxb which uses the names of the parameters.

From Rvmmin and Rcgmin


An indicator vector, having 1 for each parameter that is "free" or unconstrained, and 0 for any parameter that is fixed or MASKED for the duration of the optimization.

Note that the function bmchk() in packages optimx and optimz contains a much more extensive examination of the bounds on parameters. In particular, it considers the issues of inadmissible bounds (lower > upper), when to convert a pair of bounds
where upper["parameter"] - lower["parameter"] < tol to a fixed or masked parameter (maskadded) and whether parameters outside of bounds should be moved to the nearest bound (parchanged). It may be useful to use inadmissible to refer to situations where a lower bound is higher than an upper bound and infeasible where a parameter is outside the bounds.

From optimr/optimrx

The function optimr() can call many different "optimizers" (actually function minimization methods that may include bounds and possibly masks). These may be specified by setting the lower and upper bounds equal for the parameters to be fixed. This seems a simple method for specifying masks, but does pose some issues. For example, what happens when the upper bound is only very slightly greater than the lower bound. Also should we stop or declare an error if starting values are NOT on the fixed value.

Of these methods, my preference is now to use the last one -- setting lower and upper bounds equal, and furthermore setting the starting value to this fixed value, and otherwise declaring an error. The approach does not add any special argument for masking, and is relatively obvious to novice users. However, such users may be tempted to put in narrow bounds rather than explicit equalities, and this could have deleterious consequences.

Internal structures

bdmsk is the internal structure used in Rcgmin and Rvmmin to handle bounds constraints as well as masks. There is one element of bdmsk for each parameter, and in Rcgmin and Rvmmin, this is used on input to specify parameter i as fixed or masked by setting bdmsk[i] <- 0. Free parameters have their bdmsk element 1, but during optimization in the presence of bounds, we can set other values. The full set is as follows

Not all these possibilities will be used by all methods that use bdmsk.

The -1 and -3 are historical, and arose in the development of BASIC codes for Nash J C and Walker-Smith M (1987) Nonlinear parameter estimation: and integrated system in BASIC New York: Dekker. Now available for free download from archive.org. (\url{https://archive.org/details/NLPE87plus}). In particular, adding 2 gives 1 for an upper bound and -1 for a lower bound, simplifying the expression to decide if an optimization trial step will move away from a bound.

Proposed approaches

Because masks (fixed parameters) reduce the dimensionality of the optimization problem, we can consider modifying the problem to the lower dimension space. This is Duncan Murdoch's suggestion, using

The major advantage of this approach is explicit dimension reduction. The main disadvantage is the effort of transformation at every step of an optimization.

An alternative is to use the bdmsk vector to mask the optimization search or adjustment vector, including gradients and (approximate) Hessian matrices. A 0 element of bdmsk "multiplies" any adjustment. The principal difficulty is to ensure we do not essentially divide by zero in applying any inverse Hessian. This approach avoids forward, inverse and fn1. However, it may hide the reduction in dimension, and caution is necessary in using the function and its derived gradient, Hessian and derived information.

Examples of use

?? put them in

#   cat("Fv=",f," at ")
#   print(x)
sq.g <- function(x){
   gg<- 2*(x - yy)
xx <- c(.3, 4)
uncans <- Rvmmin(xx, sq, sq.g)
mybm <- c(0,1) # fix parameter 1
cans <- Rvmmin(xx, sq, sq.g, bdmsk=mybm)
weed <- c(5.308, 7.24, 9.638, 12.866, 17.069, 23.192, 31.443,
   38.558, 50.156, 62.948, 75.995, 91.972)
ii <- 1:12
wdf <- data.frame(weed, ii)
weedux <- nlxb(weed~b1/(1+b2*exp(-b3*ii)), start=c(b1=200, b2=50, b3=0.3)) 
weedcx <- nlxb(weed~b1/(1+b2*exp(-b3*ii)), start=c(b1=200, b2=50, b3=0.3), masked=c("b1")) 
rfn <- function(bvec, weed=weed, ii=ii){
  res <- rep(NA, length(ii))
  for (i in ii){
    res[i]<- bvec[1]/(1+bvec[2]*exp(-bvec[3]*i))-weed[i]
weeduf <- nlfb(start=c(200, 50, 0.3),resfn=rfn,weed=weed, ii=ii)
weedcf <- nlfb(start=c(200, 50, 0.3),resfn=rfn,weed=weed, ii=ii, maskidx=c(1))

Try the nlsr package in your browser

Any scripts or data that you put into this service are public.

nlsr documentation built on Oct. 6, 2017, 3 a.m.