memory usage

[AlexandreWadoux]

It happened several times that I had a memory usage problem with optimUSER, even though my user-defined objective function was very fast to compute and was not taking any memory. I finally managed to make a reproducible example.

I define my objective function as the spatially average prediction error variance obtained by kriging.

See the example below:

library(sp)
library(raster)
library(gstat)
data(meuse.grid, package = "sp")

# create the candidate locations based on a fine resolution grid
candi <- disaggregate(rasterFromXYZ(meuse.grid), fact=3)
candi <- as.data.frame(candi, xy=T, na.rm=T )

# study area boundaries
data(meuse.area, package = 'sp')
meuse.area <- list(Polygons(list(Polygon(meuse.area)), "meuse.area"))
meuse.area <- SpatialPolygons(meuse.area)

# make a grid
coordinates(meuse.grid) <- ~x+y
gridded(meuse.grid) <- TRUE

# fit a geostatistical model using log(zinc) as variable of interest
data(meuse, package='sp')
coordinates(meuse) <- ~x+y
lzn.vgm = variogram(log(zinc)~1, meuse)
lzn.fit = fit.variogram(lzn.vgm, model = vgm(1, "Sph", 900, 1))

FUN <- function(points, meuse.grid, model ){
  
  # take the matrix and change to dataframe, give a value to zinc (required by gstat, 
  # not used to compute the error variance)
  points <- as.data.frame(points)
  meuse <- points
  meuse$zinc <- 100
  
  # compute kriging prediction error variance
  coordinates(meuse)<- ~x+y
  lzn.kriged = krige(log(zinc)~1, meuse, meuse.grid, model = lzn.fit)
  varpred <- lzn.kriged$var1.var
  
  # avoid possible NA, return the averaged prediction error variance
  if(is.na(mean(varpred))){return(Inf)}else{return(mean(varpred))}
  
}
  
schedule <- scheduleSPSANN(chains = 100, initial.temperature = 10, 
                           temperature.decrease = 0.5, 
                           x.min = 0, y.min = 0, 
                           cellsize = 0)

set.seed(2001)
res <- optimUSER(points = 10, candi = candi[,c('x', 'y')], fun=FUN,  
                  schedule = schedule, track=T, plotit = T, boundary = meuse.area, 
                 meuse.grid= meuse.grid, model = lzn.fit)

it returns:
Error: cannot allocate vector of size 5.8 Gb
Very surprising!

It disappears by either:
1 - reducing the number of possible candidate locations, for example candi = meuse.grid
2 - setting y.max and x.max. I read in the documentation that, by default, x.max and y.max are set to half the maximum distance in the x- and y-coordinates, respectively. This means in the example above that x.max = (max(candi$x) - min(candi$x))/2 and y.max =(max(candi$y) - min(candi$y))/2. I find x.max = 1553 and y.max = 2073. When I set x.max and y.max it now works without problem, without even the need to decrease the number of candidate locations.