Bugzilla – Bug 14071

bug in '...' of constrOptim

Last modified: 2010-04-11 16:53:45 UTC

From: Elizabeth Purdom <epurdom@stat.berkeley.edu> Dear all, There appears to be a bug in how constrOptim handles ... arguments that are suppose to be passed to optim, according to the documentation. This means you can't get the hessian to be returned, for example (so this is a real problem, and not just a question of mistaken documentation). Looking at the code, it appears that a call to the user-defined f includes the ..., when the ... should only be passed to the optim call. I get around it by putting a dummy 'hessian=TRUE' argument in my f function, but this is not how the function should work. I show the relevant problem in the code, give an example, and also give my sessionInfo() call below. Thanks, Elizabeth ##Copy of the relevant segment of the code of constrOptim and where I think the problem might be: for (i in 1L:outer.iterations) { obj.old <- obj r.old <- r theta.old <- theta fun <- function(theta, ...) { ##this one's okay R(theta, theta.old, ...)##this one's okay } gradient <- function(theta, ...) {##this one's okay dR(theta, theta.old, ...)##this one's okay } a <- optim(theta.old, fun, gradient, control = control, method = method, ...)##this one's okay r <- a$value if (is.finite(r) && is.finite(r.old) && abs(r - r.old)/(outer.eps + abs(r - r.old)) < outer.eps) break theta <- a$par obj <- f(theta, ...)##this one's NOT okay if (obj > obj.old) break } ###Here is an example modified from the examples of the help page of constrOptim: > fr <- function(x) { ## Rosenbrock Banana function + x1 <- x[1] + x2 <- x[2] + 100 * (x2 - x1 * x1)^2 + (1 - x1)^2 + } > grr <- function(x) { ## Gradient of 'fr' + x1 <- x[1] + x2 <- x[2] + c(-400 * x1 * (x2 - x1 * x1) - 2 * (1 - x1), + 200 * (x2 - x1 * x1)) + } > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), ci=c(-1,-1)) $par [1] 0.9999761 0.9999522 $value [1] 5.708626e-10 $counts function gradient 6 1 $convergence [1] 0 $message NULL $outer.iterations [1] 14 $barrier.value [1] -0.0001999198 > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), ci=c(-1,-1),hessian=TRUE) Error in f(theta, ...) : unused argument(s) (hessian = TRUE) #add in a dummy argument to my f function > fr <- function(x,hessian=TRUE) { ## Rosenbrock Banana function + x1 <- x[1] + x2 <- x[2] + 100 * (x2 - x1 * x1)^2 + (1 - x1)^2 + } > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), ci=c(-1,-1),hessian=TRUE) $par [1] 0.9999761 0.9999522 $value [1] 5.708626e-10 $counts function gradient 6 1 $convergence [1] 0 $message NULL $hessian [,1] [,2] [1,] 801.9599 -399.9905 [2,] -399.9905 199.9952 $outer.iterations [1] 14 $barrier.value [1] -0.0001999198 ###My Session info > sessionInfo() R version 2.10.0 (2009-10-26) x86_64-apple-darwin9.8.0 locale: [1] en_US.UTF-8/en_US.UTF-8/C/C/en_US.UTF-8/en_US.UTF-8 attached base packages: [1] stats graphics grDevices utils datasets methods base other attached packages: [1] GEOquery_2.10.0 RCurl_1.2-1 bitops_1.0-4.1 Biobase_2.5.8 biomaRt_2.1.0 projectManager_1.0 lattice_0.17-26 RColorBrewer_1.0-2 [9] XML_2.6-0 loaded via a namespace (and not attached): [1] grid_2.10.0 tools_2.10.0 -- Elizabeth Purdom Assistant Professor Department of Statistics UC, Berkeley Evans Hall, Rm 433 epurdom@stat.berkeley.edu (510) 642-6154 (office) (510) 642-7892 (fax)

From: "Ravi Varadhan" <RVaradhan@jhmi.edu> PARTS: 2 Hi Elizabeth, I have fixed this problem. See the attached function. There is also another bug that can create a problem when control$fnscale = -1, i.e. when the objective function is to be maximized. I have commented this in the code and have also fixed this problem. There is a word of caution. Hessian is problematic when the converged parameter estimate is on the boundary of the feasible region, since the derivatives are not defined there. Also, note that the default method is Nelder-Mead. If you want to use "BFGS", you have to specify the gradient function. I have written an extension of constrOptim that can use numerical gradients, and can also handle nonlinear constraints. Hope this helps, Ravi. ---------------------------------------------------------------------------- ------- Ravi Varadhan, Ph.D. Assistant Professor, The Center on Aging and Health Division of Geriatric Medicine and Gerontology Johns Hopkins University Ph: (410) 502-2619 Fax: (410) 614-9625 Email: rvaradhan@jhmi.edu Webpage: http://www.jhsph.edu/agingandhealth/People/Faculty_personal_pages/Varadhan.h tml ---------------------------------------------------------------------------- -------- -----Original Message----- From: r-devel-bounces@r-project.org [mailto:r-devel-bounces@r-project.org] On Behalf Of epurdom@stat.berkeley.edu Sent: Wednesday, November 18, 2009 1:10 PM To: r-devel@stat.math.ethz.ch Cc: R-bugs@r-project.org Subject: [Rd] bug in '...' of constrOptim (PR#14071) Dear all, There appears to be a bug in how constrOptim handles ... arguments that are suppose to be passed to optim, according to the documentation. This means you can't get the hessian to be returned, for example (so this is a real problem, and not just a question of mistaken documentation). Looking at the code, it appears that a call to the user-defined f includes the ..., when the ... should only be passed to the optim call. I get around it by putting a dummy 'hessian=TRUE' argument in my f function, but this is not how the function should work. I show the relevant problem in the code, give an example, and also give my sessionInfo() call below. Thanks, Elizabeth ##Copy of the relevant segment of the code of constrOptim and where I think the problem might be: for (i in 1L:outer.iterations) { obj.old <- obj r.old <- r theta.old <- theta fun <- function(theta, ...) { ##this one's okay R(theta, theta.old, ...)##this one's okay } gradient <- function(theta, ...) {##this one's okay dR(theta, theta.old, ...)##this one's okay } a <- optim(theta.old, fun, gradient, control = control, method = method, ...)##this one's okay r <- a$value if (is.finite(r) && is.finite(r.old) && abs(r - r.old)/(outer.eps + abs(r - r.old)) < outer.eps) break theta <- a$par obj <- f(theta, ...)##this one's NOT okay if (obj > obj.old) break } ###Here is an example modified from the examples of the help page of constrOptim: > fr <- function(x) { ## Rosenbrock Banana function + x1 <- x[1] + x2 <- x[2] + 100 * (x2 - x1 * x1)^2 + (1 - x1)^2 + } > grr <- function(x) { ## Gradient of 'fr' + x1 <- x[1] + x2 <- x[2] + c(-400 * x1 * (x2 - x1 * x1) - 2 * (1 - x1), + 200 * (x2 - x1 * x1)) + } > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), ci=c(-1,-1)) $par [1] 0.9999761 0.9999522 $value [1] 5.708626e-10 $counts function gradient 6 1 $convergence [1] 0 $message NULL $outer.iterations [1] 14 $barrier.value [1] -0.0001999198 > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), ci=c(-1,-1),hessian=TRUE) Error in f(theta, ...) : unused argument(s) (hessian = TRUE) #add in a dummy argument to my f function > fr <- function(x,hessian=TRUE) { ## Rosenbrock Banana function + x1 <- x[1] + x2 <- x[2] + 100 * (x2 - x1 * x1)^2 + (1 - x1)^2 + } > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), ci=c(-1,-1),hessian=TRUE) $par [1] 0.9999761 0.9999522 $value [1] 5.708626e-10 $counts function gradient 6 1 $convergence [1] 0 $message NULL $hessian [,1] [,2] [1,] 801.9599 -399.9905 [2,] -399.9905 199.9952 $outer.iterations [1] 14 $barrier.value [1] -0.0001999198 ###My Session info > sessionInfo() R version 2.10.0 (2009-10-26) x86_64-apple-darwin9.8.0 locale: [1] en_US.UTF-8/en_US.UTF-8/C/C/en_US.UTF-8/en_US.UTF-8 attached base packages: [1] stats graphics grDevices utils datasets methods base other attached packages: [1] GEOquery_2.10.0 RCurl_1.2-1 bitops_1.0-4.1 Biobase_2.5.8 biomaRt_2.1.0 projectManager_1.0 lattice_0.17-26 RColorBrewer_1.0-2 [9] XML_2.6-0 loaded via a namespace (and not attached): [1] grid_2.10.0 tools_2.10.0 -- Elizabeth Purdom Assistant Professor Department of Statistics UC, Berkeley Evans Hall, Rm 433 epurdom@stat.berkeley.edu (510) 642-6154 (office) (510) 642-7892 (fax) ______________________________________________ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel (Attached 'constrOptim2.txt' of type 'text/plain') **END

Created attachment 966 [details]

From: Elizabeth Purdom <epurdom@stat.berkeley.edu> Hi Ravi, I just briefly looked at the code (and haven't run it). But it seems like there is still the underlying problem that '...' argument is sent to both optim and f. The documentation says that ... arguments are sent to optim, so at least based on this the ... should only appear in optim. This seems prefered since you don't need an explicit argument 'hessian' added; after all the user gets to define f, so the flexibility of ... there is minimal compared to being able to pass arguments to optim. Here's the relevant code from what you sent: a <- optim(theta.old, fun, gradient, control = control, method = method, hessian=hessian, ...) r <- a$value if (is.finite(r) && is.finite(r.old) && abs(r - r.old)/(outer.eps + abs(r - r.old)) < outer.eps) break theta <- a$par obj <- f(theta, ...) ###I don't think this should have the ... Best, Elizabeth Ravi Varadhan wrote: > Hi Elizabeth, > > I have fixed this problem. See the attached function. There is also > another bug that can create a problem when control$fnscale = -1, i.e. when > the objective function is to be maximized. I have commented this in the > code and have also fixed this problem. > > There is a word of caution. Hessian is problematic when the converged > parameter estimate is on the boundary of the feasible region, since the > derivatives are not defined there. > > Also, note that the default method is Nelder-Mead. If you want to use > "BFGS", you have to specify the gradient function. I have written an > extension of constrOptim that can use numerical gradients, and can also > handle nonlinear constraints. > > > Hope this helps, > > Ravi. > > ---------------------------------------------------------------------------- > ------- > > Ravi Varadhan, Ph.D. > > Assistant Professor, The Center on Aging and Health > > Division of Geriatric Medicine and Gerontology > > Johns Hopkins University > > Ph: (410) 502-2619 > > Fax: (410) 614-9625 > > Email: rvaradhan@jhmi.edu > > Webpage: > http://www.jhsph.edu/agingandhealth/People/Faculty_personal_pages/Varadhan.h > tml > > > > ---------------------------------------------------------------------------- > -------- > > > -----Original Message----- > From: r-devel-bounces@r-project.org [mailto:r-devel-bounces@r-project.org] > On Behalf Of epurdom@stat.berkeley.edu > Sent: Wednesday, November 18, 2009 1:10 PM > To: r-devel@stat.math.ethz.ch > Cc: R-bugs@r-project.org > Subject: [Rd] bug in '...' of constrOptim (PR#14071) > > Dear all, > > There appears to be a bug in how constrOptim handles ... arguments that > are suppose to be passed to optim, according to the documentation. This > means you can't get the hessian to be returned, for example (so this is > a real problem, and not just a question of mistaken documentation). > > Looking at the code, it appears that a call to the user-defined f > includes the ..., when the ... should only be passed to the optim call. > I get around it by putting a dummy 'hessian=TRUE' argument in my f > function, but this is not how the function should work. > > I show the relevant problem in the code, give an example, and also give > my sessionInfo() call below. > > Thanks, > Elizabeth > > ##Copy of the relevant segment of the code of constrOptim and where I > think the problem might be: > for (i in 1L:outer.iterations) { > obj.old <- obj > r.old <- r > theta.old <- theta > fun <- function(theta, ...) { ##this one's okay > R(theta, theta.old, ...)##this one's okay > } > gradient <- function(theta, ...) {##this one's okay > dR(theta, theta.old, ...)##this one's okay > } > a <- optim(theta.old, fun, gradient, control = control, > method = method, ...)##this one's okay > r <- a$value > if (is.finite(r) && is.finite(r.old) && abs(r - > r.old)/(outer.eps + > abs(r - r.old)) < outer.eps) > break > theta <- a$par > obj <- f(theta, ...)##this one's NOT okay > if (obj > obj.old) > break > } > > ###Here is an example modified from the examples of the help page of > constrOptim: > > fr <- function(x) { ## Rosenbrock Banana function > + x1 <- x[1] > + x2 <- x[2] > + 100 * (x2 - x1 * x1)^2 + (1 - x1)^2 > + } > > grr <- function(x) { ## Gradient of 'fr' > + x1 <- x[1] > + x2 <- x[2] > + c(-400 * x1 * (x2 - x1 * x1) - 2 * (1 - x1), > + 200 * (x2 - x1 * x1)) > + } > > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), ci=c(-1,-1)) > $par > [1] 0.9999761 0.9999522 > > $value > [1] 5.708626e-10 > > $counts > function gradient > 6 1 > > $convergence > [1] 0 > > $message > NULL > > $outer.iterations > [1] 14 > > $barrier.value > [1] -0.0001999198 > > > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), > ci=c(-1,-1),hessian=TRUE) > Error in f(theta, ...) : unused argument(s) (hessian = TRUE) > #add in a dummy argument to my f function > > fr <- function(x,hessian=TRUE) { ## Rosenbrock Banana function > + x1 <- x[1] > + x2 <- x[2] > + 100 * (x2 - x1 * x1)^2 + (1 - x1)^2 > + } > > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), > ci=c(-1,-1),hessian=TRUE) > $par > [1] 0.9999761 0.9999522 > > $value > [1] 5.708626e-10 > > $counts > function gradient > 6 1 > > $convergence > [1] 0 > > $message > NULL > > $hessian > [,1] [,2] > [1,] 801.9599 -399.9905 > [2,] -399.9905 199.9952 > > $outer.iterations > [1] 14 > > $barrier.value > [1] -0.0001999198 > ###My Session info > > sessionInfo() > R version 2.10.0 (2009-10-26) > x86_64-apple-darwin9.8.0 > > locale: > [1] en_US.UTF-8/en_US.UTF-8/C/C/en_US.UTF-8/en_US.UTF-8 > > attached base packages: > [1] stats graphics grDevices utils datasets methods base > > other attached packages: > [1] GEOquery_2.10.0 RCurl_1.2-1 bitops_1.0-4.1 > Biobase_2.5.8 biomaRt_2.1.0 projectManager_1.0 lattice_0.17-26 > RColorBrewer_1.0-2 > [9] XML_2.6-0 > > loaded via a namespace (and not attached): > [1] grid_2.10.0 tools_2.10.0 > -- Elizabeth Purdom Assistant Professor Department of Statistics UC, Berkeley Evans Hall, Rm 433 epurdom@stat.berkeley.edu (510) 642-6154 (office) (510) 642-7892 (fax)

From: "Ravi Varadhan" <RVaradhan@jhmi.edu> Elizabeth, If you look at the arguments in optim(), you will see that the only additional argument that is in optim but not in constrOptim is the `hessian'. So, it makes sense to add that explicitly in constrOptim() as well. The other additional arguments are passed to optim via the `control' argument. So, apart from `control' and `hessian' you cannot pass any other argument to optim. Ravi. ---------------------------------------------------------------------------- ------- Ravi Varadhan, Ph.D. Assistant Professor, The Center on Aging and Health Division of Geriatric Medicine and Gerontology Johns Hopkins University Ph: (410) 502-2619 Fax: (410) 614-9625 Email: rvaradhan@jhmi.edu Webpage: http://www.jhsph.edu/agingandhealth/People/Faculty_personal_pages/Varadhan.h tml ---------------------------------------------------------------------------- -------- -----Original Message----- From: Elizabeth Purdom [mailto:epurdom@stat.berkeley.edu] Sent: Wednesday, November 18, 2009 5:12 PM To: Ravi Varadhan Cc: r-devel@stat.math.ethz.ch; R-bugs@r-project.org Subject: Re: [Rd] bug in '...' of constrOptim (PR#14071) Hi Ravi, I just briefly looked at the code (and haven't run it). But it seems like there is still the underlying problem that '...' argument is sent to both optim and f. The documentation says that ... arguments are sent to optim, so at least based on this the ... should only appear in optim. This seems prefered since you don't need an explicit argument 'hessian' added; after all the user gets to define f, so the flexibility of ... there is minimal compared to being able to pass arguments to optim. Here's the relevant code from what you sent: a <- optim(theta.old, fun, gradient, control = control, method = method, hessian=hessian, ...) r <- a$value if (is.finite(r) && is.finite(r.old) && abs(r - r.old)/(outer.eps + abs(r - r.old)) < outer.eps) break theta <- a$par obj <- f(theta, ...) ###I don't think this should have the ... Best, Elizabeth Ravi Varadhan wrote: > Hi Elizabeth, > > I have fixed this problem. See the attached function. There is also > another bug that can create a problem when control$fnscale = -1, i.e. when > the objective function is to be maximized. I have commented this in the > code and have also fixed this problem. > > There is a word of caution. Hessian is problematic when the converged > parameter estimate is on the boundary of the feasible region, since the > derivatives are not defined there. > > Also, note that the default method is Nelder-Mead. If you want to use > "BFGS", you have to specify the gradient function. I have written an > extension of constrOptim that can use numerical gradients, and can also > handle nonlinear constraints. > > > Hope this helps, > > Ravi. > > ---------------------------------------------------------------------------- > ------- > > Ravi Varadhan, Ph.D. > > Assistant Professor, The Center on Aging and Health > > Division of Geriatric Medicine and Gerontology > > Johns Hopkins University > > Ph: (410) 502-2619 > > Fax: (410) 614-9625 > > Email: rvaradhan@jhmi.edu > > Webpage: > http://www.jhsph.edu/agingandhealth/People/Faculty_personal_pages/Varadhan.h > tml > > > > ---------------------------------------------------------------------------- > -------- > > > -----Original Message----- > From: r-devel-bounces@r-project.org [mailto:r-devel-bounces@r-project.org] > On Behalf Of epurdom@stat.berkeley.edu > Sent: Wednesday, November 18, 2009 1:10 PM > To: r-devel@stat.math.ethz.ch > Cc: R-bugs@r-project.org > Subject: [Rd] bug in '...' of constrOptim (PR#14071) > > Dear all, > > There appears to be a bug in how constrOptim handles ... arguments that > are suppose to be passed to optim, according to the documentation. This > means you can't get the hessian to be returned, for example (so this is > a real problem, and not just a question of mistaken documentation). > > Looking at the code, it appears that a call to the user-defined f > includes the ..., when the ... should only be passed to the optim call. > I get around it by putting a dummy 'hessian=TRUE' argument in my f > function, but this is not how the function should work. > > I show the relevant problem in the code, give an example, and also give > my sessionInfo() call below. > > Thanks, > Elizabeth > > ##Copy of the relevant segment of the code of constrOptim and where I > think the problem might be: > for (i in 1L:outer.iterations) { > obj.old <- obj > r.old <- r > theta.old <- theta > fun <- function(theta, ...) { ##this one's okay > R(theta, theta.old, ...)##this one's okay > } > gradient <- function(theta, ...) {##this one's okay > dR(theta, theta.old, ...)##this one's okay > } > a <- optim(theta.old, fun, gradient, control = control, > method = method, ...)##this one's okay > r <- a$value > if (is.finite(r) && is.finite(r.old) && abs(r - > r.old)/(outer.eps + > abs(r - r.old)) < outer.eps) > break > theta <- a$par > obj <- f(theta, ...)##this one's NOT okay > if (obj > obj.old) > break > } > > ###Here is an example modified from the examples of the help page of > constrOptim: > > fr <- function(x) { ## Rosenbrock Banana function > + x1 <- x[1] > + x2 <- x[2] > + 100 * (x2 - x1 * x1)^2 + (1 - x1)^2 > + } > > grr <- function(x) { ## Gradient of 'fr' > + x1 <- x[1] > + x2 <- x[2] > + c(-400 * x1 * (x2 - x1 * x1) - 2 * (1 - x1), > + 200 * (x2 - x1 * x1)) > + } > > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), ci=c(-1,-1)) > $par > [1] 0.9999761 0.9999522 > > $value > [1] 5.708626e-10 > > $counts > function gradient > 6 1 > > $convergence > [1] 0 > > $message > NULL > > $outer.iterations > [1] 14 > > $barrier.value > [1] -0.0001999198 > > > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), > ci=c(-1,-1),hessian=TRUE) > Error in f(theta, ...) : unused argument(s) (hessian = TRUE) > #add in a dummy argument to my f function > > fr <- function(x,hessian=TRUE) { ## Rosenbrock Banana function > + x1 <- x[1] > + x2 <- x[2] > + 100 * (x2 - x1 * x1)^2 + (1 - x1)^2 > + } > > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), > ci=c(-1,-1),hessian=TRUE) > $par > [1] 0.9999761 0.9999522 > > $value > [1] 5.708626e-10 > > $counts > function gradient > 6 1 > > $convergence > [1] 0 > > $message > NULL > > $hessian > [,1] [,2] > [1,] 801.9599 -399.9905 > [2,] -399.9905 199.9952 > > $outer.iterations > [1] 14 > > $barrier.value > [1] -0.0001999198 > ###My Session info > > sessionInfo() > R version 2.10.0 (2009-10-26) > x86_64-apple-darwin9.8.0 > > locale: > [1] en_US.UTF-8/en_US.UTF-8/C/C/en_US.UTF-8/en_US.UTF-8 > > attached base packages: > [1] stats graphics grDevices utils datasets methods base > > other attached packages: > [1] GEOquery_2.10.0 RCurl_1.2-1 bitops_1.0-4.1 > Biobase_2.5.8 biomaRt_2.1.0 projectManager_1.0 lattice_0.17-26 > RColorBrewer_1.0-2 > [9] XML_2.6-0 > > loaded via a namespace (and not attached): > [1] grid_2.10.0 tools_2.10.0 > -- Elizabeth Purdom Assistant Professor Department of Statistics UC, Berkeley Evans Hall, Rm 433 epurdom@stat.berkeley.edu (510) 642-6154 (office) (510) 642-7892 (fax)

From: "Colin Millar" <C.Millar@MARLAB.AC.UK> Hi, i would tend to disagree, the ... arguments in optim come directly after the fn and grad functions as they are intended as extra named arguments to be passed to the fn and grad during optimisation. In constrOptim the fn and grad functions wrapped in a function to make the contraints work, so this wrapper nessarily will need to recieve the ... which is then passed on down to the orgonal user supplied fn with possible extra arguements. if the there is issue with the ... could you put the following lines instead? if (length(formals(f)) == 1) { f(theta) - mu * bar } else { f(theta, ...) - mu * bar } ? ________________________________ From: r-devel-bounces@r-project.org on behalf of Elizabeth Purdom Sent: Wed 18/11/2009 22:12 To: Ravi Varadhan Cc: R-bugs@r-project.org; r-devel@stat.math.ethz.ch Subject: Re: [Rd] bug in '...' of constrOptim (PR#14071) Hi Ravi, I just briefly looked at the code (and haven't run it). But it seems like there is still the underlying problem that '...' argument is sent to both optim and f. The documentation says that ... arguments are sent to optim, so at least based on this the ... should only appear in optim. This seems prefered since you don't need an explicit argument 'hessian' added; after all the user gets to define f, so the flexibility of ... there is minimal compared to being able to pass arguments to optim. Here's the relevant code from what you sent: a <- optim(theta.old, fun, gradient, control = control, method = method, hessian=hessian, ...) r <- a$value if (is.finite(r) && is.finite(r.old) && abs(r - r.old)/(outer.eps + abs(r - r.old)) < outer.eps) break theta <- a$par obj <- f(theta, ...) ###I don't think this should have the ... Best, Elizabeth Ravi Varadhan wrote: > Hi Elizabeth, > > I have fixed this problem. See the attached function. There is also > another bug that can create a problem when control$fnscale = -1, i.e. when > the objective function is to be maximized. I have commented this in the > code and have also fixed this problem. > > There is a word of caution. Hessian is problematic when the converged > parameter estimate is on the boundary of the feasible region, since the > derivatives are not defined there. > > Also, note that the default method is Nelder-Mead. If you want to use > "BFGS", you have to specify the gradient function. I have written an > extension of constrOptim that can use numerical gradients, and can also > handle nonlinear constraints. > > > Hope this helps, > > Ravi. > > ---------------------------------------------------------------------------- > ------- > > Ravi Varadhan, Ph.D. > > Assistant Professor, The Center on Aging and Health > > Division of Geriatric Medicine and Gerontology > > Johns Hopkins University > > Ph: (410) 502-2619 > > Fax: (410) 614-9625 > > Email: rvaradhan@jhmi.edu > > Webpage: > http://www.jhsph.edu/agingandhealth/People/Faculty_personal_pages/Varadhan.h > tml > > > > ---------------------------------------------------------------------------- > -------- > > > -----Original Message----- > From: r-devel-bounces@r-project.org [mailto:r-devel-bounces@r-project.org] > On Behalf Of epurdom@stat.berkeley.edu > Sent: Wednesday, November 18, 2009 1:10 PM > To: r-devel@stat.math.ethz.ch > Cc: R-bugs@r-project.org > Subject: [Rd] bug in '...' of constrOptim (PR#14071) > > Dear all, > > There appears to be a bug in how constrOptim handles ... arguments that > are suppose to be passed to optim, according to the documentation. This > means you can't get the hessian to be returned, for example (so this is > a real problem, and not just a question of mistaken documentation). > > Looking at the code, it appears that a call to the user-defined f > includes the ..., when the ... should only be passed to the optim call. > I get around it by putting a dummy 'hessian=TRUE' argument in my f > function, but this is not how the function should work. > > I show the relevant problem in the code, give an example, and also give > my sessionInfo() call below. > > Thanks, > Elizabeth > > ##Copy of the relevant segment of the code of constrOptim and where I > think the problem might be: > for (i in 1L:outer.iterations) { > obj.old <- obj > r.old <- r > theta.old <- theta > fun <- function(theta, ...) { ##this one's okay > R(theta, theta.old, ...)##this one's okay > } > gradient <- function(theta, ...) {##this one's okay > dR(theta, theta.old, ...)##this one's okay > } > a <- optim(theta.old, fun, gradient, control = control, > method = method, ...)##this one's okay > r <- a$value > if (is.finite(r) && is.finite(r.old) && abs(r - > r.old)/(outer.eps + > abs(r - r.old)) < outer.eps) > break > theta <- a$par > obj <- f(theta, ...)##this one's NOT okay > if (obj > obj.old) > break > } > > ###Here is an example modified from the examples of the help page of > constrOptim: > > fr <- function(x) { ## Rosenbrock Banana function > + x1 <- x[1] > + x2 <- x[2] > + 100 * (x2 - x1 * x1)^2 + (1 - x1)^2 > + } > > grr <- function(x) { ## Gradient of 'fr' > + x1 <- x[1] > + x2 <- x[2] > + c(-400 * x1 * (x2 - x1 * x1) - 2 * (1 - x1), > + 200 * (x2 - x1 * x1)) > + } > > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), ci=c(-1,-1)) > $par > [1] 0.9999761 0.9999522 > > $value > [1] 5.708626e-10 > > $counts > function gradient > 6 1 > > $convergence > [1] 0 > > $message > NULL > > $outer.iterations > [1] 14 > > $barrier.value > [1] -0.0001999198 > > > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), > ci=c(-1,-1),hessian=TRUE) > Error in f(theta, ...) : unused argument(s) (hessian = TRUE) > #add in a dummy argument to my f function > > fr <- function(x,hessian=TRUE) { ## Rosenbrock Banana function > + x1 <- x[1] > + x2 <- x[2] > + 100 * (x2 - x1 * x1)^2 + (1 - x1)^2 > + } > > constrOptim(c(-1.2,0.9), fr, grr, ui=rbind(c(-1,0),c(0,-1)), > ci=c(-1,-1),hessian=TRUE) > $par > [1] 0.9999761 0.9999522 > > $value > [1] 5.708626e-10 > > $counts > function gradient > 6 1 > > $convergence > [1] 0 > > $message > NULL > > $hessian > [,1] [,2] > [1,] 801.9599 -399.9905 > [2,] -399.9905 199.9952 > > $outer.iterations > [1] 14 > > $barrier.value > [1] -0.0001999198 > ###My Session info > > sessionInfo() > R version 2.10.0 (2009-10-26) > x86_64-apple-darwin9.8.0 > > locale: > [1] en_US.UTF-8/en_US.UTF-8/C/C/en_US.UTF-8/en_US.UTF-8 > > attached base packages: > [1] stats graphics grDevices utils datasets methods base > > other attached packages: > [1] GEOquery_2.10.0 RCurl_1.2-1 bitops_1.0-4.1 > Biobase_2.5.8 biomaRt_2.1.0 projectManager_1.0 lattice_0.17-26 > RColorBrewer_1.0-2 > [9] XML_2.6-0 > > loaded via a namespace (and not attached): > [1] grid_2.10.0 tools_2.10.0 > -- Elizabeth Purdom Assistant Professor Department of Statistics UC, Berkeley Evans Hall, Rm 433 epurdom@stat.berkeley.edu (510) 642-6154 (office) (510) 642-7892 (fax) ______________________________________________ R-devel@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-devel ______________________________________________________________________ This email has been scanned by the MessageLabs Email Security System. For more information please visit http://www.messagelabs.com/email ______________________________________________________________________ [[alternative HTML version deleted]]

NOTES: Seems user confusion

Audit (from Jitterbug): Sat Nov 21 08:34:02 2009 ripley changed notes

argument 'hessian' added to constrOptim for 2.12.0