SPOT (version 2.1.10)

spot: spot

Description

Sequential Parameter Optimization. This is one of the main interfaces for using the SPOT package. Based on a user-given objective function and configuration, spot finds the parameter setting that yields the lowest objective value (minimization). To that end, it uses methods from the fields of design of experiment, statistical modeling / machine learning and optimization.

Usage

spot(x = NULL, fun, lower, upper, control = list(), ...)

Arguments

x

is an optional start point (or set of start points), specified as a matrix. One row for each point, and one column for each optimized parameter.

fun

is the objective function. It should receive a matrix x and return a matrix y. In case the function uses external code and is noisy, an additional seed parameter may be used, see the control$seedFun argument below for details.

lower

is a vector that defines the lower boundary of search space. This determines also the dimensionality of the problem.

upper

is a vector that defines the upper boundary of search space.

control

is a list with control settings for spot. See spotControl.

...

additional parameters passed to fun.

Value

This function returns a list with:

xbest

Parameters of the best found solution (matrix).

ybest

Objective function value of the best found solution (matrix).

x

Archive of all evaluation parameters (matrix).

y

Archive of the respective objective function values (matrix).

count

Number of performed objective function evaluations.

msg

Message specifying the reason of termination.

modelFit

The fit of the last build model, i.e., an object returned by the last call to the function specified by control$model.

Examples

# NOT RUN {
## Most simple example: Kriging + LHS + predicted
## mean optimization (not expected improvement)
res <- spot(,funSphere,c(-2,-3),c(1,2),control=list(funEvals=15))
res$xbest
## With expected improvement
res <- spot(,funSphere,c(-2,-3),c(1,2),
   control=list(funEvals=15,modelControl=list(target="ei")))
res$xbest
### With additional start point:
#res <- spot(matrix(c(0.05,0.1),1,2),funSphere,c(-2,-3),c(1,2))
#res$xbest
#res <- spot(,funSphere,c(-2,-3),c(1,2),
#    control=list(funEvals=50))
#res$xbest
### Use local optimization instead of LHS
#res <- spot(,funSphere,c(-2,-3),c(1,2),
#    control=list(optimizer=optimLBFGSB))
#res$xbest
### Random Forest instead of Kriging
#res <- spot(,funSphere,c(-2,-3),c(1,2),
#    control=list(model=buildRandomForest))
#res$xbest
### LM instead of Kriging
#res <- spot(,funSphere,c(-2,-3),c(1,2),
#    control=list(model=buildLM)) #lm as surrogate
#res$xbest
### LM and local optimizer (which for this simple example is perfect)
#res <- spot(,funSphere,c(-2,-3),c(1,2),
#    control=list(model=buildLM, optimizer=optimLBFGSB))
#res$xbest
### Lasso and local optimizer NLOPTR
#res <- spot(,funSphere,c(-2,-3),c(1,2), 
#    control=list(direct= TRUE, model=buildLasso, optimizer = optimNLOPTR))
#res$xbest
### Kriging and local optimizer LBFGSB 
#res <- spot(,funSphere,c(-2,-3),c(1,2), 
#    control=list(direct= TRUE, model=buildKriging, optimizer = optimLBFGSB))
#res$xbest
### Kriging and local optimizer NLOPTR 
#res <- spot(,funSphere,c(-2,-3),c(1,2), 
#    control=list(direct= TRUE, model=buildKriging, optimizer = optimNLOPTR))
#res$xbest
### Or a different Kriging model:
#res <- spot(,funSphere,c(-2,-3),c(1,2),
#    control=list(model=buildKrigingDACE, optimizer=optimLBFGSB))
#res$xbest
## With noise: (this takes some time)
#res1 <- spot(,function(x)funSphere(x)+rnorm(nrow(x)),c(-2,-3),c(1,2),
#		control=list(funEvals=100,noise=TRUE)) #noisy objective
#res2 <- spot(,function(x)funSphere(x)+rnorm(nrow(x)),c(-2,-3),c(1,2),
#		control=list(funEvals=100,noise=TRUE,replicates=2,
#		designControl=list(replicates=2))) #noise with replicated evaluations
#res3 <- spot(,function(x)funSphere(x)+rnorm(nrow(x)),c(-2,-3),c(1,2),
#		control=list(funEvals=100,noise=TRUE,replicates=2,OCBA=T,OCBABudget=1,
#		designControl=list(replicates=2))) #and with OCBA
### Check results with non-noisy function:
#funSphere(res1$xbest)
#funSphere(res2$xbest)
#funSphere(res3$xbest)
## The following is for demonstration only, to be used for random number
## seed handling in case of external noisy target functions.
#res3 <- spot(,function(x,seed){set.seed(seed);funSphere(x)+rnorm(nrow(x))},
#    c(-2,-3),c(1,2),control=list(funEvals=100,noise=TRUE,seedFun=1))
##
## Next Example: Handling factor variables
## Note: factors should be coded as integer values, i.e., 1,2,...,n
## create a test function:
braninFunctionFactor <- function (x) {
  y <- (x[2]  - 5.1/(4 * pi^2) * (x[1] ^2) + 5/pi * x[1]  - 6)^2 +
    10 * (1 - 1/(8 * pi)) * cos(x[1] ) + 10
  if(x[3]==1)
    y <- y +1
  else if(x[3]==2)
    y <- y -1
  y
}
## vectorize
objFun <- function(x){apply(x,1,braninFunctionFactor)}
set.seed(1)
res <- spot(fun=objFun,lower=c(-5,0,1),upper=c(10,15,3),
    control=list(model=buildKriging,
      types= c("numeric","numeric","factor"),
      optimizer=optimLHD))
res$xbest
res$ybest
# }