optimx is an R package that extends and enhances the optim() function of base R, in particular by unifying the call to many solvers. This Gitlab project has been established to document optimx and its development. A more detailed history of the package is given at the bottom of this document.
From within R, the command
install.packages("optimx")will install the optimx from the CRAN repository https://cran.r-project.org/
of which a copy is here on github.com/nashjc/. Thus if the devtools package is installed in R on your computer, you can issue the commands
library("devtools")
install_github("nashjc/optimx")The test version of optimx on R-forge in the optimizer project (https://r-forge.r-project.org/R/?group_id=395) is now deprecated.
optimr() allows any of the solvers linked to the package to be called via the call
result <- optimr(par, fn, gr, hess, method=“a_solver”, lower=lo, upper=up)
where - par is a vector of the initial parameter values for the function - fn is the objective function to be minimized - gr is a function to compute the gradient of fn (or NULL if none is provided) - hess is a function to compute the hessian of fn (or NULL if gr=NULL or none is provided) - method provides the name of the solver, here given a token name - lower and upper provide bounds on the parameters for solvers that can use them
See the man page for other arguments that allow various controls or extra features.
opm() allows multiple solvers to be called via a single statement so that their performance may be compared. It is INEFFICIENT, WASTEFUL, and strongly DISCOURAGED to use opm() as a general-purpose optimization tool. The call is similar to that of optimr() but method is now a vector of names of solvers. See the man page for more details.
Usage:
result<-opm(par, fn, gr, hess, lower=lo, upper=up, method=c(“solver1”, “solver2”, “solver3”))
with arguments as for optimr() except for method, which is now a vector of character names.
The package contains several solver functions. See specific man pages for details. Some of these are intended to be called via optimr() and NOT independently.
The table includes methods from other packages or base-R as well as solvers in optimx.
| solver | source | gradient | hessian | bounds | algorithm class | Notes |
|---|---|---|---|---|---|---|
| anms | pracma | polytope | ||||
| BFGS | optim() | optional | variable metric | |||
| bobyqa | minqa | yes | model & descend | |||
| CG | optim() | optional | conjugate grad | Deprecated | ||
| hjkb | dfoptim | yes | Hooke&Jeeves | |||
| hjn | optimx | yes | Hooke&Jeeves | |||
| lbfgs | lbfgs | optional | LM quasi Newton | |||
| L-BFGS-B | optim() | optional | yes | LM quasi Newton | ||
| lbfgsb3c | lbfgsb3c | optional | yes | LM quasi Newton | ||
| mla | marqLevAlg | optional | optional | Newton-like | ||
| ncg | optimx | required | yes | conjugate grad | ||
| Nelder-Mead | optim() | polytope | ||||
| newuoa | minqa | model & descend | ||||
| nlm | base-R | optional | optional | Newton-like | ||
| nlminb | base-R | optional | optional | yes | multiple | |
| nlnm | nloptr | yes | polytope | |||
| nmkb | dfoptim | yes | polytope | |||
| nvm | optimx | required | yes | variable metric | ||
| pracmanm | pracma | polytope | ||||
| Rcgmin | optimx | required | yes | conjugate grad | legacy compatibility | |
| Rtnmin | optimx | required | yes | trunc. Newton | ||
| Rvmmin | optimx | required | yes | variable metric | legacy compatibility | |
| slsqp | nloptr | optional | yes | quad programming | ||
| snewtm | optimx | required | required | Newton-like | Didactic | |
| snewton | optimx | required | required | Newton-like | Didactic | |
| spg | BB | optional | yes | proj. gradient | ||
| subplex | subplex | polytope | ||||
| tnewt | nloptr | optional | yes | trunc. Newton | ||
| ucminf | ucminf | optional | variable metric | |||
| uobyqa | minqa | model & descend | Use newuoa | |||
| ————— | ———– | ———- | ———- | ——- | —————— | ———————— |
Note that there are other versions of some of these algorithms in different packages, e.g., package nloptr has a bobyqa() function.
It is relatively easy to modify optimr.R and ctrldefault.R to add a new solver. See the vignette “Using and extending the optimx package” (file Extend-optimx.pdf).
The following files provide tools to support the use of the optimx package and other optimization tasks. See the man pages for details on usage.
Some solvers, in particular ncg() and nvm() can explicitly deal with lower and upper bounds constraints that are set equal for some parameters. Such fixed parameters can be useful when a parameter is not generally altered to try to optimize the function, but may be allowed into the optimization in the future. Such constraints are sometimes called “masks”. The initial parameter value must be consistent with the bounds settings.
Use of equal lower and upper bounds may be possible for other solvers that handle bounds constraints, but users are urged to test this possibility before trusting results.
The optimx package was first placed on CRAN in 2009 by John Nash and Ravi Varadhan. The purpose was to allow a single syntax to call one or several different optimization solvers with a single optimization problem. The package was quite successful, and the optimx() function within the package is still present in the current package with some minor corrections. However, the structure of the optimx() function made it difficult to extend and maintain. Moreover, the output was slightly altered from that of the base-R optim() function.
The optimr and optimrx (experimental) packages were an attempt to resolve the maintenance issue and to address some deficiencies in optimx(). In particular, the optimr() function was introduced to mirror the base-R optim() function but with more solvers. The multiple-solver feature of the optimx() function was replaced with opm(), while polyalgorithm and multiple initial parameter sets (starts) were handled with polyopt() and multistart() respectively, avoiding some complicated option possibilities in optimx(). Moreover, the optimr() function now permits all solvers to use parameter scaling via the control element (actually a vector) parscale.
In 2016/2017 it became clear that there were a number of optimization tools that could be consolidated into a single packages. The current optimx package replaces:
It also includes two new Newton-like solvers, a didactic Hooke and Jeeves solver, and a simple interface to various approximate gradient routines. As and when it is sensible to do so, the subsumed packages will be archived.
Note that some development versions of packages remain on R-forge. Use of these is at your own risk.