tests/sim/20220314-sim/look-at-results.R

In jlivsey/Dvar: What the Package Does (One Line, Title Case)

library(tidyverse, warn.conflicts = FALSE, quietly = TRUE)
dim0 <- c(2, 2, 7, 3, 2, 43)

simDir <- "~/Github/Dvar/tests/sim/20220314-sim/bottom-up///"
nsim <- 20

# ---- Load Results ----
simLevels <- 2:12

R <- matrix(NA, 7224, length(simLevels))
R <- array(NA, dim = c(7224, length(simLevels), nsim))
colnames(R) <- letters[1:length(simLevels)]
timeRunMat <- matrix(NA, nsim, length(simLevels))

for(jj in seq_along(simLevels)){
  simNum_here <- simLevels[jj]
  print(simNum_here)
  fileName <- sprintf("results%d.RData", simNum_here)
  print(fileName)
  load(file.path(simDir, fileName))
  R[, jj, ] <- t(coefEsts)
  simColNames <- paste0("sim", simNum_here)
  colnames(R)[jj] <- simColNames
  timeRunMat[, jj] <- timeRun
}


# Reformat timeRunMat (some records are in mins and some are in hours)
timeRunMat[timeRunMat > 10] <- timeRunMat[timeRunMat > 10] / 60

boxplot(timeRunMat)



head(R)


# ---- True Data ----
load(file.path(simDir, "true-table.Rdata")) # Loads array A
Av <- c(A)

# ---- Absolue Error from truth ----
absErr <- abs(R - Av)
par(mfrow = c(1, 2), mar = c(3, 3, 3, .1))
absErrMat <- aperm(absErr, c(1, 3, 2)) %>% matrix(ncol = 10)
boxplot(absErrMat, main = "absolute error")
boxplot(absErrMat, ylim = c(0, 9), main = "zoom in")


# ---- Histograms ----
hist(absErrMat[, 1], breaks = 50)
hist(absErrMat[, 10], breaks = 50)


# ---- summary ----
View(apply(R, 2, summary))

qprob <- function(x){sum(x[x < quantile(x, prob=.95)])}

apply(absErr, 2:3, sum) %>% t() %>% boxplot()


# ---- Absolute Error of margins in workload ----

M <- matrix(NA, length(marPack), 7224)
for(i in 1:length(marPack)) {
  # extract i-th element from marPack and formulate as full vectors not 0.
  mar <- marginZero2vec(marPack[[i]], mydim = dim0)
  # Find next row for design matrix
  Zv <- c(arraySetOnes(dim0, mar))
  M[i, ] <- Zv
}
# M is the margin extraction matrix. M %*% datavector will return margin

marTru <- M %*% Av

marEst <- array(NA, dim = c(length(marPack), length(simLevels), nsim))
for(i in seq_along(simLevels)){
  for(j in seq_len(nsim)){
    marEst[, i, j] <- M %*% R[, i, j]
  }
}

marAbsErr <- abs(c(marTru) - marEst)
boxplot(marAbsErr[, , 1])
boxplot(marAbsErr, ylim = c(0, 20), main = "zoom in to (0, 1)")

# Stack replicates into same column as workload
for(i in seq_len(nsim)){
  if(i == 1){
    marAbsErrStacked = marAbsErr[, , 1]
    next
  }
  marAbsErrStacked <- rbind(marAbsErrStacked, marAbsErr[, , i])
}

# ---- Put results into table ----

maxErr  = apply(marAbsErrStacked, 2, max)
meanErr = apply(marAbsErrStacked, 2, mean)
q3Err   = apply(marAbsErrStacked, 2, function(x){quantile(x, .75)})

outTab <- rbind(maxErr, meanErr, q3Err)
colnames(outTab) = paste("WL", 1:length(simLevels), sep = "-")
xtable::xtable(outTab)


# ---- Comparison to laplace variance ----
geoMod   <- .3
queryMod <- .3
b        <- sqrt(2) / 5 / geoMod / queryMod

2 * b^2
apply(c(marTru) - marEst, 2, var)

save(R, A, y_true, noise,
     file = file.path(simDir, "20220818-bottomUp-forEric-simdate20220314.RData"))