Hi all,
I'm using optimx (version 2013.8.7) to perform parameter estimation with
the nelder-mead method, and received a warning that the parameters are on
different scales, which can hurt optimization performance on derivative
free methods. This warning is accurate as some parameters are small
(between 0 and 1) and others can be quite large.
So, I tried using the parscale option in the control list to put them on
more equal scaling. There are 8 parameters, so I supplied a vector of
length 8 to the parscale option, which if I understand correctly, would
scale the large parameters down by their starting values so everything will
start out near 1 Here is the call to optimx itself I used:
fit <- optimx(par = c(ER=.6,LR=.035,TR
=.05,FR=.1,alpha=50,lambda=.4,Tmin=.5,Tmax=40),
fn = RT_ErrorFcn,
method = "Nelder-Mead"
control = list(maxit=1000,
parscale = c(1,1,1,1,50, 1,1,40)),
fcn = model$fn, # passed to RT_ErrorFcn
fix = model$fixed, #RT_ErrorFcn
obs = data) # passed to RT_ErrorFcn
However, the warning about different parameter scales is still given. Is it
intended behavior for the warning to be given even when the parscale option
is used? Or am I misunderstanding the point of the parscale option, or
implementing it wrong?
Thanks for any info on this topic.
- Will
[[alternative HTML version deleted]]