max_surnew_robinv: Optimization of the surnew criterion

In IRSN/RobustInv: "Robust inversion of expensive black-box functions"

Description

Minimizes the surnew criterion using either discrete or genetic optimization

Usage

max_surnew_robinv(lower, upper, optimcontrol = NULL, opt.index,
  inv.index, integration.param, T, model, new.noise.var = NULL)

Arguments

lower	Array of size d. Lower bound of the input domain.
upper	Array of size d. Upper bound of the input domain.
optimcontrol	A list with the following fields. method : can be either equal to "discrete" for discrete optimization or "genoud" (default) for optimization using the genoud package. (I) If optimcontrol$method="discrete" the user can set the field optim.points which is a matrix with d columns containing all the points tested to find the optimum. If not set we use 100*d random points generated uniformly. (II) If optimcontrol$method="genoud" the user can set the arguments of the genoud algorithm. pop.size: population size of each generation (default: 50*d), max.generation: maximum number of generations used to find the optimum (default: 2*d), wait.generation: the algorithm stops if no improvement is done for wait.generation generations (default: 1). Other parameters of the genoud algorithm can be set, namely BFGSburnin, parinit, unif.seed, int.seed, P1, P2, ..., until P9. See the documentation of the genoud function.
opt.index	Array with integers corresponding to the indices of the nuisance parameters.
inv.index	Array with integers corresponding to the indices of the controlled parameters.
integration.param	An object obtained from a call to the integration_design_robinv.
T	Target threshold.
model	The current kriging model. km object.
new.noise.var	Noise variance of the new observations. Leave to NULL for noiseless functions. For noisy functions, any non zero value is valid and will give the same result.

Value

A list with the following fields. (i) par: optimizer of the sampling criterion. (ii) value: minimum of the sampling criterion. (iii) allvalues: when a discrete optimization is performed, this contains all the values of the criterion for the tested points.

Author(s)

Clement Chevalier [e-mail]

Examples

library(KrigInv)
myfun <- branin_robinv
d <- 3

set.seed(8)

n0 <- 30
T <- 10
opt.index <- c(3)
inv.index <- c(1,2)
lower <- rep(0,times=d)
upper <- rep(1,times=d)

design <- matrix(runif(d*n0),nrow=n0)
response <- myfun(design)
model <- km(formula = ~1,design = design,response = response,covtype = "matern3_2")

integcontrol <- list(distrib = "surnew",n.points = 20,finaldistrib="surnew",
                     n.candidates=50,nsimu=1000,n.optpoints = 50,
                     choose_optpoints=TRUE,n.optpoints.candidates=500)
## Not run: 
obj <- integration_design_robinv(integcontrol = integcontrol,d=d,lower=lower,upper=upper,
                                 opt.index=opt.index,inv.index=inv.index,model=model,T=T)

# one try with discrete optimization:
optimcontrol <- list(method="discrete")

result <- max_surnew_robinv(lower = lower,upper = upper,optimcontrol = optimcontrol,
                            opt.index = opt.index,inv.index = inv.index,
                            integration.param = obj,T = T,model = model)

result$par
result$value

# one try with genoud optimization:
optimcontrol <- list(method="pso",pop.size = 200,max.generations=3)

result2 <- max_surnew_robinv(lower = lower,upper = upper,optimcontrol = optimcontrol,
                             opt.index = opt.index,inv.index = inv.index,
                             integration.param = obj,T = T,model = model)

result2$par
result2$value

## End(Not run)

IRSN/RobustInv documentation built on Nov. 20, 2019, 10:46 p.m.