SPOTVignetteNutshell

Introduction

The performance of modern search heuristics relies crucially on their parameterizations—or, statistically speaking, on their factor settings.
Finding good parameter settings for an optimization algorithm will be referred to as tuning.
The version of SPOT presented in this article is implemented in R.
It can be used as an algorithm tuner or directly as an optimizer.
This vignette is an abbreviated version of a the technical report “In a Nutshell-The Sequential Parameter Optimization Toolbox”.
It illustrates how spot can be called.
The extended version of this vignette is available on arXiv and will be updated regularly.

Sequential Parameter Optimization Examples: How to Call SPOT

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
#>            [,1]      [,2]
#> [1,] -0.1086201 0.1184503

With expected improvement

res <- spot(,funSphere,c(-2,-3),c(1,2),
    control=list(funEvals=15,modelControl=list(target="ei")))
res$xbest
#>            [,1]      [,2]
#> [1,] -0.1086201 0.1184503

With additional start point:

res <- spot(matrix(c(0.05,0.1),1,2),funSphere,c(-2,-3),c(1,2))
res$xbest
#>             [,1]        [,2]
#> [1,] -0.06104759 -0.05040567

Larger budget:

res <- spot(,funSphere,c(-2,-3),c(1,2),
    control=list(funEvals=50))
res$xbest
#>            [,1]        [,2]
#> [1,] 0.02088315 -0.03783177

Use local optimization instead of LHS

res <- spot(,funSphere,c(-2,-3),c(1,2),
   control=list(optimizer=optimLBFGSB))
 res$xbest
#>             [,1]         [,2]
#> [1,] -0.01573496 -0.008860252

Random Forest instead of Kriging

res <- spot(,funSphere,c(-2,-3),c(1,2),
     control=list(model=buildRandomForest))
res$xbest
#>           [,1]      [,2]
#> [1,] 0.1531584 0.3294388

LM instead of Kriging

res <- spot(,funSphere,c(-2,-3),c(1,2),
     control=list(model=buildLM)) #lm as surrogate
res$xbest
#>           [,1]      [,2]
#> [1,] 0.1531584 0.3294388

Bayesian Optimization

res <- spot(,funSphere,c(-2,-3),c(1,2),
     control=list(model=buildBO)) #BO as surrogate
res$xbest
#>              [,1]        [,2]
#> [1,] -0.009619762 -0.01032117

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
#>             [,1]       [,2]
#> [1,] -0.02651579 -0.1091904

Lasso and local optimizer NLOPTR

res <- spot(,funSphere,c(-2,-3),c(1,2), 
   control=list(model=buildLasso, optimizer = optimNLOPTR))
#> Warning: Option grouped=FALSE enforced in cv.glmnet, since < 3 observations per
#> fold

#> Warning: Option grouped=FALSE enforced in cv.glmnet, since < 3 observations per
#> fold

#> Warning: Option grouped=FALSE enforced in cv.glmnet, since < 3 observations per
#> fold

#> Warning: Option grouped=FALSE enforced in cv.glmnet, since < 3 observations per
#> fold

#> Warning: Option grouped=FALSE enforced in cv.glmnet, since < 3 observations per
#> fold

#> Warning: Option grouped=FALSE enforced in cv.glmnet, since < 3 observations per
#> fold

#> Warning: Option grouped=FALSE enforced in cv.glmnet, since < 3 observations per
#> fold

#> Warning: Option grouped=FALSE enforced in cv.glmnet, since < 3 observations per
#> fold

#> Warning: Option grouped=FALSE enforced in cv.glmnet, since < 3 observations per
#> fold

#> Warning: Option grouped=FALSE enforced in cv.glmnet, since < 3 observations per
#> fold
res$xbest
#>           [,1]      [,2]
#> [1,] 0.1531584 0.3294388

Kriging and local optimizer LBFGSB

res <- spot(,funSphere,c(-2,-3),c(1,2), 
   control=list(model=buildKriging, optimizer = optimLBFGSB))
res$xbest
#>             [,1]         [,2]
#> [1,] -0.01573496 -0.008860252

Kriging and local optimizer NLOPTR

res <- spot(,funSphere,c(-2,-3),c(1,2), 
     control=list(model=buildKriging, optimizer = optimNLOPTR))
res$xbest
#>             [,1]         [,2]
#> [1,] -0.01440329 0.0006858711

Or a different Kriging model:

res <- spot(,funSphere,c(-2,-3),c(1,2),
 control=list(model=buildKrigingDACE, optimizer=optimLBFGSB))
res$xbest
#>             [,1]         [,2]
#> [1,] -0.03586733 -0.004407048

With noise: (this takes some time)

# noisy objective
res1 <- spot(,function(x)funSphere(x)+rnorm(nrow(x)),c(-2,-3),c(1,2),
        control=list(funEvals=40,noise=TRUE)) 
# noise with replicated evaluations
res2 <- spot(,function(x)funSphere(x)+rnorm(nrow(x)),c(-2,-3),c(1,2),
        control=list(funEvals=40,noise=TRUE,replicates=2,
        designControl=list(replicates=2))) 
# and with OCBA
res3 <- spot(,function(x)funSphere(x)+rnorm(nrow(x)),c(-2,-3),c(1,2),
        control=list(funEvals=40,noise=TRUE,replicates=2,OCBA=TRUE,OCBABudget=1,
        designControl=list(replicates=2))) 
# Check results with non-noisy function:
funSphere(res1$xbest)
#>           [,1]
#> [1,] 0.9811919
funSphere(res2$xbest)
#>         [,1]
#> [1,] 1.74417
funSphere(res3$xbest)
#>          [,1]
#> [1,] 1.084404

Random number seed handling

The following is for demonstration only, to be used for random number seed handling in case of external noisy target functions.

res1a <- spot(,function(x,seed){set.seed(seed);funSphere(x)+rnorm(nrow(x))},
     c(-2,-3),c(1,2),control=list(funEvals=25,noise=TRUE,seedFun=1))

res1b <- spot(,function(x,seed){set.seed(seed);funSphere(x)+rnorm(nrow(x))},
     c(-2,-3),c(1,2),control=list(funEvals=25,noise=TRUE,seedFun=1))

res2 <- spot(,function(x,seed){set.seed(seed);funSphere(x)+rnorm(nrow(x))},
     c(-2,-3),c(1,2),control=list(funEvals=25,noise=TRUE,seedFun=2))

sprintf("Should be equal: %f = %f. Should be different:  %f", res1a$ybest, res1b$ybest, res2$ybest)
#> [1] "Should be equal: -1.296329 = -1.296329. Should be different:  -0.889934"

Handling factor variables

Note: factors should be coded as integer values, i.e., 1,2,…,n First, we create a test function with a factor variable:

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
   return(y)
}

Vectorize the test function.

objFun <- function(x){apply(x,1,braninFunctionFactor)}

Run spot.

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$xbest
#>          [,1]     [,2] [,3]
#> [1,] 2.619386 2.725642    2
 res$ybest
#>           [,1]
#> [1,] 0.6777176

High dimensional problem

n <- 10
a <- rep(0,n)
b <- rep(1,n)

First, we consider the default spot setting with buildKriging().

tic <- proc.time()[3]
res0 <- spot(x=NULL, funSphere, lower = a, upper = b, 
             control=list(funEvals=30))
toc <- proc.time()[3]
sprintf("value: %f, time: %f",  res0$ybest, toc-tic)
#> [1] "value: 0.332121, time: 7.336000"

Then, we use the buildGaussianProcess() model.

tic <- proc.time()[3]
res1 <-  spot(x=NULL, funSphere, lower = a, upper = b, 
             control=list(funEvals=30, 
                          model = buildGaussianProcess))
toc <- proc.time()[3]
sprintf("value: %f, time: %f",  res1$ybest, toc-tic)
#> [1] "value: 0.530886, time: 0.684000"

Run SPOT with logging

## run spot without log
res <- spot(fun = funSphere,
            lower=c(0,0),
            upper=c(100,100)
)
## run spot with log
funSphereLog <- function(x){
  cbind(funSphere(x),x)
}
res2 <- spot(fun = funSphereLog,
            lower=c(0,0),
            upper=c(100,100)
)
res$logInfo
#> [1] NA
res2$logInfo
#>            [,1]       [,2]
#>  [1,]  8.648076 81.4784571
#>  [2,] 71.771945 66.5887761
#>  [3,] 44.933187 41.8506996
#>  [4,] 14.297134 39.5437814
#>  [5,] 25.642638 58.9784849
#>  [6,] 56.561623 79.4369705
#>  [7,] 69.785541 27.2369075
#>  [8,] 92.321611 93.7035707
#>  [9,] 32.408116  7.8101754
#> [10,] 87.968361 10.1114951
#> [11,]  3.152969  3.1352284
#> [12,] 35.979045 83.1545726
#> [13,]  1.401563  0.3142767
#> [14,] 32.165836 18.4191631
#> [15,] 27.534938 83.0152530
#> [16,] 34.745003 77.4768640
#> [17,] 21.418092 95.5431781
#> [18,] 42.504694 61.7054862
#> [19,] 87.533618 15.9641532
#> [20,] 74.720086 84.2995194

Hybrid optimization

res <- spot(fun = funSphere, lower = c(-5,-5),
                upper = c(5,5), 
                control = list(funEvals = 20,
                directOpt = optimNLOPTR,
                directOptControl = list(funEvals = 10)
                ))
str(res)
#> List of 9
#>  $ xbest   : num [1, 1:2] 0 0
#>  $ ybest   : num 0
#>  $ x       : num [1:33, 1:2] -4.135 2.177 -0.507 -3.57 -2.436 ...
#>  $ y       : num [1:33, 1] 27.009 7.492 0.921 13.84 6.739 ...
#>  $ logInfo : logi NA
#>  $ count   : int 20
#>  $ msg     : chr "budget exhausted"
#>  $ modelFit:List of 33
#>   ..$ thetaLower      : num 1e-04
#>   ..$ thetaUpper      : num 100
#>   ..$ types           : chr [1:2] "numeric" "numeric"
#>   ..$ algTheta        :function (x = NULL, fun, lower, upper, control = list(), ...)  
#>   ..$ budgetAlgTheta  : num 200
#>   ..$ optimizeP       : logi FALSE
#>   ..$ useLambda       : logi TRUE
#>   ..$ lambdaLower     : num -6
#>   ..$ lambdaUpper     : num 0
#>   ..$ startTheta      : NULL
#>   ..$ reinterpolate   : logi TRUE
#>   ..$ target          : chr "y"
#>   ..$ modelInitialized: logi TRUE
#>   ..$ x               : num [1:19, 1:2] -4.135 2.177 -0.507 -3.57 -2.436 ...
#>   ..$ y               : num [1:19, 1] 27.009 7.492 0.921 13.84 6.739 ...
#>   ..$ normalizeymin   : num 0
#>   ..$ normalizeymax   : num 1
#>   ..$ scaledx         : num [1:19, 1:2] 0 0.7544 0.4337 0.0675 0.2031 ...
#>   ..$ normalizexmin   : num [1:2] -4.14 -4.22
#>   ..$ normalizexmax   : num [1:2] 4.23 4.55
#>   ..$ dmodeltheta     : num [1:2] 0.122 0.141
#>   ..$ Lambda          : num -5.96
#>   ..$ dmodellambda    : num 1.09e-06
#>   ..$ Theta           : num [1:2] -0.915 -0.852
#>   ..$ yonemu          : num [1:19, 1] -323 -342 -349 -336 -343 ...
#>   ..$ ssq             : num 14020
#>   ..$ mu              : num 350
#>   ..$ Psi             : num [1:19, 1:19] 1 0.929 0.95 0.968 0.986 ...
#>   ..$ Psinv           : num [1:19, 1:19] 77206 -36217 35306 -27605 -92449 ...
#>   ..$ nevals          : num 1200
#>   ..$ like            : num [1, 1] 1.96
#>   ..$ returnCrossCor  : logi FALSE
#>   ..$ min             : num 0.0988
#>   ..- attr(*, "class")= chr "kriging"
#>  $ ybestVec: num [1:20] 0.921 0.921 0.921 0.921 0.921 ...

Handling constraints

Definitions and constraints

library(babsim.hospital)
n <- 29 
reps <- 2
funEvals <- 3*n 
size <- 2*n
x0 <- matrix(as.numeric(babsim.hospital::getParaSet(5374)[1,-1]),1,)
bounds <- getBounds()
a <- bounds$lower
b <- bounds$upper
g <- function(x) {
      return(rbind(a[1] - x[1], x[1] - b[1], a[2] - x[2], x[2] - b[2], 
                   a[3] - x[3], x[3] - b[3], a[4] - x[4], x[4] - b[4], 
                   a[5] - x[5], x[5] - b[5], a[6] - x[6], x[6] - b[6], 
                   a[7] - x[7], x[7] - b[7], a[8] - x[8], x[8] - b[8], 
                   a[9] - x[9], x[9] - b[9], a[10] - x[10], x[10] - b[10],
                   a[11] - x[11], x[11] - b[11], a[12] - x[12],  x[12] - b[12],
                   a[13] - x[13], x[13] - b[13], a[14] - x[14],  x[14] - b[14],
                   a[15] - x[15], x[15] - b[15], a[16] - x[16],  x[16] - b[16],
                   a[17] - x[17], x[17] - b[17], a[18] - x[18],  x[18] - b[18],
                   a[19] - x[19], x[19] - b[19], a[20] - x[20],  x[20] - b[20],
                   a[21] - x[21], x[21] - b[21], a[22] - x[22],  x[22] - b[22],
                   a[23] - x[23], x[23] - b[23], a[24] - x[24],  x[24] - b[24],
                   a[25] - x[25], x[25] - b[25], a[26] - x[26],  x[26] - b[26],
                   a[27] - x[27], x[27] - b[27], x[15] + x[16] - 1, 
                   x[17] + x[18] + x[19] - 1, x[20] + x[21] - 1, x[23] + x[29] - 1)
      )
  }

res <- spot(
  x = x0,
  fun = funBaBSimHospital,
  lower = a,
  upper = b,
  verbosity = 0,
  control = list(
    funEvals = 2 * funEvals,
    noise = TRUE,
    designControl = list(# inequalityConstraint = g,
      size = size,
      retries = 1000),
    optimizer = optimNLOPTR,
    optimizerControl = list(
      opts = list(algorithm = "NLOPT_GN_ISRES"),
      eval_g_ineq = g
    ),
    model =  buildKriging,
    plots = FALSE,
    progress = TRUE,
    directOpt = optimNLOPTR,
    directOptControl = list(funEvals = 0),
    eval_g_ineq = g
  )
)
print(res)

GECCO Industrial Challenge 2021

A description of the challenge can be found here: GECCO Industrial Challenge 2021. In short the goal of the challenge is to find an optimal parameter configuration for the BabSim.Hospital simulator. This is a noisy and complex real-world problem.

Evaluation Using the Docker Container

In order to be able to execute the necessary code of the GECCO Industrial challenge 2021 you will need to have Docker installed in your machine. On your terminal console an evaluation of the BabSim.Hospital should looks like the command below. This command will automatically download the Docker image with the BabSim.Hospital code in it (may need sudo rights to download). Take care, the formatting of the symbols - and ’ can cause this command not to work on your terminal:

# docker run --rm mrebolle/r-geccoc:Track1 -c 'Rscript objfun.R "6,7,3,3,3,5,3,3,25,17,2,1,0.25,0.05,0.07,0.005,0.07,1e-04,0.08,0.25,0.08,0.5,1e-06,2,1e-06,1e-06,1,2,0.5"'

An optimization run with SPOT, using the Docker command call as objective function, can be directly implemented in R as follows:

library(SPOT)

evalFun <- function(candidateSolution){
    evalCommand <- paste0("docker run --rm mrebolle/r-geccoc:Track1 -c ", "'","Rscript objfun.R ")
    parsedCandidate <- paste(candidateSolution, sep=",", collapse = ",")
    return(as.numeric(system(paste0(evalCommand, '"', parsedCandidate, '"', "'"), intern = TRUE)))
}

#The BabSim.Hospital requires 29 parameters. Here we specify the upper and lower bounds
lower <- c(6,7,3,3,3,5,3,3,25,17,2,1,0.25,0.05,0.07,
           0.005,0.07,1e-04,0.08,0.25,0.08,0.5,1e-06,
           2,1e-06,1e-06,1,2,0.5)

upper<- c(14,13,7,9,7,9,5,7,35,25,5,7,2,0.15,0.11,0.02,
          0.13,0.002,0.12,0.35,0.12,0.9,0.01,4,1.1,0.0625,
          2,5,0.75)

wFun <- wrapFunction(evalFun)

n <- 29 
reps <- 2
funEvals <- 10*n 
size <- 2*n
x0<-matrix(lower,nrow = 1)

res <- spot(x = x0,
  fun = wFun,
  lower = lower,
  upper = upper,
  control = list(
    funEvals = 2 * funEvals,
    noise = TRUE,
    designControl = list(
      size = size,
      retries = 1000),
    optimizer = optimNLOPTR,
    optimizerControl = list(
      opts = list(algorithm = "NLOPT_GN_ISRES")
    ),
    model =  buildKriging,
    plots = TRUE,
    progress = TRUE,
    directOpt = optimNLOPTR,
    directOptControl = list(funEvals = 0)
  )
)

Evaluation Using the babsim.hospital R Package

The optimization of the BabSim.Hospital parameters can also be executed directly using the babsim.hospital package.

The babsim.hospital package can be installed by downloading the source from the Gitlab repository and building the package.

git clone http://owos.gm.fh-koeln.de:8055/bartz/babsim.hospital.git

library(SPOT)
library(babsim.hospital)

n <- 29 
reps <- 2
funEvals <- 3*n 
size <- 2*n
#Get suggested parameter values as initial point in the optimization run
x0 <- matrix(as.numeric(babsim.hospital::getParaSet(5374)[1,-1]),1,)
bounds <- getBounds()
a <- bounds$lower
b <- bounds$upper
g <- function(x) {
      return(rbind(a[1] - x[1], x[1] - b[1], a[2] - x[2], x[2] - b[2], 
                   a[3] - x[3], x[3] - b[3], a[4] - x[4], x[4] - b[4], 
                   a[5] - x[5], x[5] - b[5], a[6] - x[6], x[6] - b[6], 
                   a[7] - x[7], x[7] - b[7], a[8] - x[8], x[8] - b[8], 
                   a[9] - x[9], x[9] - b[9], a[10] - x[10], x[10] - b[10],
                   a[11] - x[11], x[11] - b[11], a[12] - x[12],  x[12] - b[12],
                   a[13] - x[13], x[13] - b[13], a[14] - x[14],  x[14] - b[14],
                   a[15] - x[15], x[15] - b[15], a[16] - x[16],  x[16] - b[16],
                   a[17] - x[17], x[17] - b[17], a[18] - x[18],  x[18] - b[18],
                   a[19] - x[19], x[19] - b[19], a[20] - x[20],  x[20] - b[20],
                   a[21] - x[21], x[21] - b[21], a[22] - x[22],  x[22] - b[22],
                   a[23] - x[23], x[23] - b[23], a[24] - x[24],  x[24] - b[24],
                   a[25] - x[25], x[25] - b[25], a[26] - x[26],  x[26] - b[26],
                   a[27] - x[27], x[27] - b[27], x[15] + x[16] - 1, 
                   x[17] + x[18] + x[19] - 1, x[20] + x[21] - 1, x[23] + x[29] - 1)
      )
  }

wrappedFunBab <- function(x){
  print(SPOT::funBaBSimHospital(x, region = 5374, nCores = 1))
}
res <- spot(
  x = x0,
  fun = wrappedFunBab,
  lower = a,
  upper = b,
  control = list(
    funEvals = 2 * funEvals,
    noise = TRUE,
    designControl = list(
      size = size,
      retries = 1000),
    optimizer = optimNLOPTR,
    optimizerControl = list(
      opts = list(algorithm = "NLOPT_GN_ISRES"),
      eval_g_ineq = g
    ),
    model =  buildKriging,
    plots = FALSE,
    progress = TRUE,
    directOpt = optimNLOPTR,
    directOptControl = list(funEvals = 0),
    eval_g_ineq = g
  )
)
print(res)