[R] optim() has a non-consistent way to send parameter for different methods

Marc Girondot marc_grt at yahoo.fr
Fri Jun 23 18:21:45 CEST 2017


When optim() is used with method="BFGS", the name of parameters within 
the vector are transmitted (see below, first example).

When method="Brent", the name of parameter (only one parameter can be 
fitted with Brent method) is not transmitted. As there is only one, of 
course, we know which parameter it is, but it makes things 
non-consistent between methods.

It would be better that same convention is used for different method.

Marc

Tested in R-3.4.0

For example, here:

@@@@@@@@@@@
Method BFGS
@@@@@@@@@@@

# The names of values in par are transmitted
fitnorm_meansd<-function(par, val) {
   print(par)
   -sum(dnorm(x=val, mean=par["mean"], sd=par["sd"], log = TRUE))
}

val <- rnorm(100, mean=20, sd=2)
p<-c(mean=20, sd=2)
result<-optim(par=p, fn=fitnorm_meansd, val=val, method="BFGS")

The print(par) shows the named vector:
 > result<-optim(par=p, fn=fitnorm_meansd, val=val, method="BFGS")
mean   sd
   20    2
   mean     sd
20.001  2.000
   mean     sd
19.999  2.000
   mean     sd
20.000  2.001
etc...

@@@@@@@@@@@
Method Brent
@@@@@@@@@@@

# The name of value in par is not transmitted

fitnorm_mean<-function(par, val) {
   print(par)
   -sum(dnorm(x=val, mean=par, sd=2, log = TRUE))
}

val <- rnorm(100, mean=20, sd=2)
p<-c(mean=20)
result<-optim(par=p, fn=fitnorm_mean, val=val, method="Brent", lower=10, 
upper=30)


The print(par) does not show named vector:
 > result<-optim(par=p, fn=fitnorm_mean, val=val, method="Brent", 
lower=10, upper=30)
[1] 17.63932
[1] 22.36068



More information about the R-help mailing list