R/ctrldefault.R

##################################################################
ctrldefault <- function(npar) { 
# THIS IS FULL VERSION FOR optimrx
#
     ## These are DEFAULTS. They may be nonsense in some contexts.

      allmeth <- c("BFGS", "CG", "Nelder-Mead", "L-BFGS-B", "nlm", "nlminb", 
                "lbfgsb3", "Rcgmin", "Rcgmin2", "Rcgdescent", "Rtnmin", "Rvmmin",
                "Rvmminq", "snewton", "snewtonm", "spg", "ucminf", "newuoa",
                "bobyqa", "nmkb", "hjkb", "hjn", "lbfgs", "subplex", "lbfgsb3c")

#  allpkg has package where element of allmeth is found
#  This list must match list allmeth in length
      allpkg <-  c("stats", "stats", "stats", "stats", "stats", "stats",
                "lbfgsb3c", "optimx", "Rcgmin2", "Rcgdescent", "optimx", "optimx",
                "optimx", "optimx", "optimx", "BB", "ucminf", "minqa",
           "minqa", "dfoptim", "dfoptim", "optimx", "lbfgs", "subplex", "lbfgsb3c")

     # 160628: uobyqa removed as it fails hobbs from 1,1,1 unscaled

      bdmeth <- c("L-BFGS-B", "nlminb", "lbfgsb3", "lbfgsb3c", "Rcgmin", "Rcgmin2", "Rtnmin",
                 "Rvmmin", "Rvmminq", "bobyqa", "nmkb", "hjkb", "hjn")

      maskmeth <- c("Rcgmin", "Rcgmin2", "Rvmmin", "Rvmminq", "hjn")
 
# offset changed from 100 to 1000 on 180710
      ctrl.default <- list(
        acctol = 0.0001, # used for acceptable point test in backtrack linesearch
        all.methods = FALSE, # we do NOT want all methods to be the default
        allmeth = allmeth, # to define the set of all methods
        allpkg = allpkg, # to list all the packages required
      	badval = (0.5)*.Machine$double.xmax, # use this value as a flag of non-computable item
        bdmeth = bdmeth, # list of bounds-constrained methods
        bigval = .Machine$double.xmax*0.01, # a very large number (note smaller than badval)
        defgrapprox = "grfwd", # use forward approximation as default. Could argue for grcentral
        defmethod = "Nelder-Mead", # same default as optim(). Do we want this?
        defstep=1, # default stepsize is 1 (Newton stepsize)
        dowarn = TRUE,  # generally want to turn on warnings
        eps = 1e-07, # a tolerance for a small quantity (about single precision level)
        epstol = .Machine$double.eps, # but this is the machine epsilon
        fnscale = 1.0, # Normally scale function by 1. -1 maximizes functions
        grtesttol=(.Machine$double.eps)^(1/3), # a test tolerance for equality for numeric and 
        # analytic gradients
        have.bounds = FALSE, # normally have UNCONSTRAINED function
        hessasymtol = 0.0001, # tolerance for testing Hessian asymmetry. Note that it 
        # cannot be too small or we get too many false positives
        hesstesttol=(.Machine$double.eps)^(1/3), # See grtesttol. This is for Hessian approx.
        keepinputpar = FALSE, # When TRUE do NOT allow bounds check to change parameter values
        kkt = TRUE, # Normally test KKT conditions. May take a LONG time.
        kkttol = 0.001, # tolerance for testing KKT small gradient condition 
        kkt2tol = 1.0E-6, # tolerance for testing KKT curvature condition
        maskmeth = maskmeth, # list of methods that allow masks (fixed parameters)
        maximize = FALSE, # normally MINIMIZE (see fnscale)
#        maxit = 500*round(sqrt(npar+1)), # limit on number of iterations or gradient evaluations
        maxit = 500, # limit on number of iterations or gradient evaluations
#        maxfeval = 5000*round(sqrt(npar+1)), # limit on function evaluations
        maxfeval = 500*round(sqrt(npar+1)), # limit on function evaluations
        offset = 1000.0, # used for equality test (a + offset) == (b + offset)
        parchanged = FALSE, # set TRUE when bounds check has changed parameter values
        parscale = rep(1, npar), # vector of scaling factors for parameters. Try to get
        # scaled parameters to have magnitude in range (1, 10)
        save.failures = TRUE,
      	scaletol = 3, 
        stepdec = 0.2, 
        stepinc = 10.0,
        steplen0 = 0.75, 
        steplenn = 1.0, 
        stepmax = 5,
        stepmin = 0,
        stepredn = 0.2,
        stopbadupdate = FALSE,
        tol = 0, 
        trace = 0,
        type = 2, # for optim::CG Polak Ribiere
        watch = FALSE,
        cgoffset = 100,
        cgstepredn = 0.15,
        cgoldstep = 0.8,
        qiilev = 0 # 
      )
}
##################################################################

dispdefault <- function(ctrl) { # Display the control vector using cat and print
  cat("Control vector (ctrl) multiple element items first:\n")
  cat("allmeth:")
  print(ctrl$allmeth)
  cat("allpkg:")
  print(ctrl$allpkg)
  cat("bdmeth:")
  print(ctrl$bdmeth)
  cat("maskmeth:")
  print(ctrl$maskmeth)
  cat("parscale:")
  print(ctrl$parscale)
  cat("acctol=",ctrl$acctol,"  all.methods=",ctrl$all.methods,"  badval=",ctrl$badval,
    "  bigval=",ctrl$bigval,"\n")
  cat("defgrapprox=",ctrl$defgrapprox,"  defmethod=",ctrl$defmethod,"  defstep=",ctrl$defstep,"\n")
  cat("dowarn=",ctrl$dowarn,"  eps=",ctrl$eps,"  epstol=",ctrl$epstol,"  fnscale=",ctrl$fnscale,"\n")
  cat("grtesttol=",ctrl$grtesttol,"  have.bounds=",ctrl$have.bounds,"  hessasymptol=",ctrl$hessasymptol,"\n")
  cat("hesttesttol=",ctrl$hesstesttol,"  keepinputpar=",ctrl$keepinputpar,"  kkt=",ctrl$kkt,"\n")
  cat("kkttol=",ctrl$kkttol,"  kkt2tol=",ctrl$kkt2tol,"  maximize=",ctrl$maximize,"\n")
  cat("maxit=",ctrl$maxit,"  maxfeval=",ctrl$maxfeval,"  offset=",ctrl$offset,
     "  parchanged=",ctrl$parchanged,"\n")
  cat("  save.failures=",ctrl$save.failures,"  scaletol=",ctrl$scaletol,"\n")
  cat("stepdec=",ctrl$stepdec,"  steplen0=",ctrl$steplen0,"  stepmax=",ctrl$stepmax,
     "  stepmin=",ctrl$stepmin,"\n")
  cat("stepredn=",ctrl$stepredn,"  stopbadupdate=",ctrl$stopbadupdate,"  tol=",ctrl$tol,"\n")
  cat("trace=",ctrl$trace,"  watch=",ctrl$watch,"\n")  
  cat("=========================================\n")  
}

Try the optimx package in your browser

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

optimx documentation built on June 14, 2019, 3:01 p.m.