mixopt {mixopt}

Mixed variable optimization using coordinate descent

Description

Mixed variable optimization using coordinate descent

Usage

mixopt(
  par,
  fn,
  gr = NULL,
  global = "multistart",
  local = "coorddesc",
  ...,
  method,
  verbose = 0,
  track
)

mixopt_blockcd(
  par,
  fn,
  gr = NULL,
  ...,
  control = list(),
  maxblocksize = NULL,
  method,
  fngr = NULL,
  maxiter = 100,
  maxeval = NULL,
  maxtime = NULL,
  verbose = 0,
  track = FALSE
)

mixopt_coorddesc(
  par,
  fn,
  gr = NULL,
  ...,
  method,
  maxiter = 100,
  maxeval = NULL,
  maxtime = NULL,
  verbose = 0,
  track = FALSE
)

mixopt_multistart(
  par,
  fn,
  gr = NULL,
  ...,
  method,
  fngr = NULL,
  n0 = 20,
  n1 = 2,
  maxiter = 100,
  maxeval = NULL,
  verbose = 0,
  groupeval = FALSE,
  track = FALSE
)

Arguments

par

List of parameters

fn

Function to evaluate

gr

Gradient of fn

global

Global optimization algorithm to use. 'FALSE' if you only want local optimization.

local

Local optimization algorithm to use.

...

Additional args

method

Optimization method

verbose

How much to print. 0 is none, 1 is standard, 2 is some, 3 is a lot, etc.

track

Should it track the parameters evaluated and value?

control

Parameters for optimizing.

maxblocksize

The maximum number of continuous dimensions that should be placed into a single block.

fngr

Function that returns the function and gradient value for the given input as a list with names "fn" and "gr".

maxiter

Maximum number of outer iterations. For coordinate descent, one iteration is a loop over each parameter.

maxeval

Maximum number of function evaluations. It may go over this number while in an inner optimization loop, but will exit after that.

maxtime

Maximum time to run in seconds. Not an exact limit, only checks occasionally.

n0

For multistart, number of random initial points to evaluate.

n1

For multistart, number of best starts to optimize with. You should have 'n0' less than 'n1', potentially by a large factor. gradient descent.

groupeval

Can multiple inputs be evaluated at once? This can speed up greatly for certain circumstances. Use "matrix" to have it give a set of points as rows of a matrix to all be evaluated at once.

Value

List

References

https://en.wikipedia.org/wiki/Coordinate_descent

https://en.wikipedia.org/wiki/Coordinate_descent

https://www.uv.es/rmarti/paper/docs/multi2.pdf

Examples

# Simple 1D example
mixopt_blockcd(par=list(mopar_cts(2,8)), fn=function(x) {(4.5-x[1])^2})
# With gradient (isn't faster)
mixopt_blockcd(par=list(mopar_cts(2,8)), fn=function(x) {(4.5-x[1])^2},
               gr=function(x) {-2*(4.5-x[1])})

# 1D discrete ordered
mixopt_blockcd(par=list(mopar_ordered(100:10000)),
                 fn=function(x) {(x[1] - 500.3)^2})

# 2D: one continuous, one factor
mixopt_blockcd(par=list(mopar_cts(2,8), mopar_unordered(letters[1:6])),
                 fn=function(x) {ifelse(x[2] == 'b', -1, 0) +
                                 (4.5-x[1])^2})
# Simple 1D example
mixopt_coorddesc(par=list(mopar_cts(2,8)), fn=function(x) {(4.5-x[1])^2})

# 1D discrete ordered
mixopt_coorddesc(par=list(mopar_ordered(100:10000)),
                 fn=function(x) {(x[1] - 500.3)^2})

# 2D: one continuous, one factor
mixopt_coorddesc(par=list(mopar_cts(2,8), mopar_unordered(letters[1:6])),
                 fn=function(x) {ifelse(x[2] == 'b', -1, 0) +
                                 (4.5-x[1])^2})
# 2D
library(ggplot2)
library(dplyr)
f6 <- function(x) {-(-x[1]*.5*sin(.5*x[1])*1 - 1e-2*x[2]^2 +
                       .2*x[1] - .3*x[2])}
ContourFunctions::cf_func(f6, xlim=c(0,100), ylim=c(-100,100))
m6 <- mixopt_coorddesc(par=list(mopar_cts(0,100), mopar_cts(-100,100)),
                       fn=f6, track = TRUE)
plot_track(m6)
ms6 <- mixopt_multistart(par=list(mopar_cts(0,100), mopar_cts(-100,100)),
                         fn=f6, track = TRUE)
plot_track(ms6)
ContourFunctions::cf_func(f6, xlim=c(0,100), ylim=c(-100,100),
                          gg = TRUE) +
  geom_point(data=as.data.frame(matrix(unlist(ms6$track$par),
                                       ncol=2, byrow=TRUE)) %>%
               bind_cols(newbest=ms6$track$newbest),
             aes(V1, V2, color=newbest), alpha=.5)