inst/tinytest/test-solving-ilp.R

In anticlust: Subset Partitioning via Anticlustering

library("anticlust")

# all levels of heuristicism work and that exact approach has best objective
conditions <- expand.grid(m = 1:4, p = 2)
for (k in 1:nrow(conditions)) {
  m_features <- conditions[k, "m"]
  p_anticlusters <- conditions[k, "p"]
  n_elements <- p_anticlusters * 5 # n must be multiplier of p
  features <- matrix(rnorm(n_elements * m_features), ncol = m_features)
  ## Traverse through levels of heuristicism
  obj_values <- rep(NA, 4)
  anti_list <- list()
  for (i in c(TRUE, FALSE)) {
    anticlusters <- anticlust:::exact_anticlustering(as.matrix(dist(features)),
                                         p_anticlusters, preclustering = i, NULL)
    anti_list[[i + 1]] <- anticlusters
    # Allow for some numeric imprecision of ILP solver:
    obj_values[i + 1]  <- round(anticlust:::diversity_objective_(anticlusters, features), 10)
  }
  ## Exact solution must have maximum objective
  expect_equal(which.max(obj_values), 1)
}

# Solving ILP works as expected for max dispersion
N <- 30
K <- 3
M <- 5
dat <- matrix(rnorm(N*M), ncol = M)

ILP <- anticlustering(dat, K = K, objective = "dispersion", method = "ilp")
HEURISTIC <- anticlustering(dat, K = K, objective = "dispersion", method = "local-maximum")

expect_true(dispersion_objective(dat, ILP) >= dispersion_objective(dat, HEURISTIC))
expect_true(all(table(ILP) == N/K))