DEoptim's own start is uniform between the bounds, which are deliberately wide (sign only). For an ELF whose feasible region is a sliver of that box, the start has no member the objective accepts, and DE never finds one: a trial is kept when it is no worse than its parent, and every value is the fail value (ELF 6.6.2, 2026-10-20: 0 of 2000 uniform draws passed the gates, with two climate terms each drawn over 0-50). This draws by rejection (drawActivePars()) and fills the rest of the population with uniform draws: DE's p-best strategy pulls the population towards its accepted members, so a few are enough to start from.

feasibleInitialPopulation(
  n,
  lower,
  upper,
  gates,
  seed,
  minFeasible = 4L,
  maxDrawsPerMember = 100L,
  cores = 1L
)

Arguments

n

integer; rows wanted. At least the population size the fit may get (clusters::clusterSetup() uses the first NP rows, NP being the number of workers).

lower, upper

the bounds, named.

gates

a function of a list of parameter sets, see spreadGates().

seed

integer; withr::with_seed(), so the population is the same on every call for an ELF.

minFeasible

integer; stop with the gate counts when fewer sets pass after maxDrawsPerMember * n draws. 4 is DEoptim's smallest population.

maxDrawsPerMember

integer; cap on the draws per member wanted.

cores

integer; forks for gates.

Value

an n x length(lower) matrix, columns named as lower, the accepted sets first; attribute feasible is how many rows were accepted.