R help - Jan 2015 - R 3.1.2 mle2() function on Windows 7 Error and multiple solutions

Hi,



I tried your problem with optimx package.  I found a better solution than that
found by mle2.



?library(optimx)



# the objective function needs to be re-written

LL2 <- function(par,y) {

lambda <- par[1]
alpha <- par[2]
beta <- par[3]
R = Nweibull(y,lambda,alpha,beta)

-sum(log(R))
}



optimx(fn=LL2,  par=c(.01,325,.8),y=y, lower=c(.00001,.00001,.00001),upper =
c(Inf, Inf,Inf),control=list(all.methods=TRUE))



# Look at the solution found by `nlminb' and `nmkb'. This is the optimal
one.  This log-likelihood is larger than that of mle2 and other optimizers in
optimx.



If this solution is not what you are looking for, your problem may be poorly
scaled.  First, make sure that the likelihood is coded correctly.  If it is
correct, then you may need to improve the scaling of the problem.





Hope this is helpful,

Ravi



	[[alternative HTML version deleted]]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

  For what it's worth, you can use either nlminb (directly) or optimx
within the mle2 wrapper by specifying the 'optimizer' parameter ...
this gives you flexibility in optimization along with the convenience
of mle2 (likelihood ratio tests via anova(), likelihood profiling, etc.)

On 15-01-15 09:26 AM, Ravi Varadhan wrote:
> Hi,
> 
> 
> 
> I tried your problem with optimx package.  I found a better
> solution than that found by mle2.
> 
> 
> 
> ?library(optimx)
> 
> 
> 
> # the objective function needs to be re-written
> 
> LL2 <- function(par,y) {
> 
> lambda <- par[1] alpha <- par[2] beta <- par[3] R = 
> Nweibull(y,lambda,alpha,beta)
> 
> -sum(log(R)) }
> 
> 
> 
> optimx(fn=LL2,  par=c(.01,325,.8),y=y, 
> lower=c(.00001,.00001,.00001),upper = c(Inf, 
> Inf,Inf),control=list(all.methods=TRUE))
> 
> 
> 
> # Look at the solution found by `nlminb' and `nmkb'. This is the 
> optimal one.  This log-likelihood is larger than that of mle2 and 
> other optimizers in optimx.
> 
> 
> 
> If this solution is not what you are looking for, your problem may
> be poorly scaled.  First, make sure that the likelihood is coded 
> correctly.  If it is correct, then you may need to improve the 
> scaling of the problem.
> 
> 
> 
> 
> 
> Hope this is helpful,
> 
> Ravi
> 
> 
> 
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)

iQEcBAEBAgAGBQJUt87VAAoJEOCV5YRblxUH9E4H/ismNjBi/diA7db1f4EtIYUz
fk0V1GIjAkhNr+gxs8bu6CBAMB2f/ufw+9ey2X6yHlzvgfwzIwNafgg9c5qVlArF
xD8A4w/4G9cRsQFX8yySEQMP7dH5tyCTeRHU0sEcTbY+vV/NtWAYpF7k36He0QnQ
Jz/Gfmjt/TTVlcsL4crr8IdOjP34mq7H1SGXKNoBymhaggkBXXjG+IlhPK3/HE4s
2LFKusdSVDiJCCR+kafwyKxk76Lf2WADw9/RaysWfW0/v5O5dWU4IuvK2//nzvts
7rKMkF9/zlT+LgLNo7LON+RTOeDtTMqyA10Vu+txQTKH4AcMP4LqYoiGMerl6O0=cHEo
-----END PGP SIGNATURE-----

A more important point that want to make is that I find few people taking
advantage of the "comparative evaluation" or benchmarking ability of
optimx.  There is no "uniformly best" optimizer for all problems. 
Different ones turn out to perform better for different problems and it is quite
difficult to know a priori which one would be the best for "my"
problem.  Thus, the benchmarking capability provided by optimx is a powerful
feature.

Ravi  
________________________________________
From: Ben Bolker <bbolker at gmail.com>
Sent: Thursday, January 15, 2015 9:29 AM
To: Ravi Varadhan; R-Help
Cc: malqurashi at KSU.EDU.SA
Subject: Re: R 3.1.2 mle2() function on Windows 7 Error and multiple solutions

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

  For what it's worth, you can use either nlminb (directly) or optimx
within the mle2 wrapper by specifying the 'optimizer' parameter ...
this gives you flexibility in optimization along with the convenience
of mle2 (likelihood ratio tests via anova(), likelihood profiling, etc.)

On 15-01-15 09:26 AM, Ravi Varadhan wrote:
> Hi,
>
>
>
> I tried your problem with optimx package.  I found a better
> solution than that found by mle2.
>
>
>
> ?library(optimx)
>
>
>
> # the objective function needs to be re-written
>
> LL2 <- function(par,y) {
>
> lambda <- par[1] alpha <- par[2] beta <- par[3] R >
Nweibull(y,lambda,alpha,beta)
>
> -sum(log(R)) }
>
>
>
> optimx(fn=LL2,  par=c(.01,325,.8),y=y,
> lower=c(.00001,.00001,.00001),upper = c(Inf,
> Inf,Inf),control=list(all.methods=TRUE))
>
>
>
> # Look at the solution found by `nlminb' and `nmkb'. This is the
> optimal one.  This log-likelihood is larger than that of mle2 and
> other optimizers in optimx.
>
>
>
> If this solution is not what you are looking for, your problem may
> be poorly scaled.  First, make sure that the likelihood is coded
> correctly.  If it is correct, then you may need to improve the
> scaling of the problem.
>
>
>
>
>
> Hope this is helpful,
>
> Ravi
>
>
>
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)

iQEcBAEBAgAGBQJUt87VAAoJEOCV5YRblxUH9E4H/ismNjBi/diA7db1f4EtIYUz
fk0V1GIjAkhNr+gxs8bu6CBAMB2f/ufw+9ey2X6yHlzvgfwzIwNafgg9c5qVlArF
xD8A4w/4G9cRsQFX8yySEQMP7dH5tyCTeRHU0sEcTbY+vV/NtWAYpF7k36He0QnQ
Jz/Gfmjt/TTVlcsL4crr8IdOjP34mq7H1SGXKNoBymhaggkBXXjG+IlhPK3/HE4s
2LFKusdSVDiJCCR+kafwyKxk76Lf2WADw9/RaysWfW0/v5O5dWU4IuvK2//nzvts
7rKMkF9/zlT+LgLNo7LON+RTOeDtTMqyA10Vu+txQTKH4AcMP4LqYoiGMerl6O0=cHEo
-----END PGP SIGNATURE-----