The repository for the "Solving the Optimal Experiment Design Problem with Mixed-Integer Convex Methods" paper.
The experiment scripts are in the ODWB folder. The models and set-ups for the different solvers can be found in the source folder in the ODWB folder.
The models and solvers tested are Boscia.jl with the classical problem formulation, SCIP.jl using the epigraph formulation, Pajarito.jl implementing the direct conic model and a custom solver based from this paper
The experiments are set up to be run on the cluster of the Zuse Institute Berlin. If you want to test instances of one dimension with a specific solver, you can use Julia's environmental variables and call the run_optimal_design.jl file directly.
using ODWB
# This sets the solver. Option are "Boscia", "SCIP", "Pajarito" and "Custom".
ENV["MODE"] = "Boscia"
# Set the criterion. Options are "A", "D" for the A-Optimal Design and D-Optimal Design, respectively.
# "AF" and "DF" are for the A-Fusion and D-Fusion.
ENV["CRITERION"] = "D"
# This decides the data type, so independent ("IND") or correlated data ("CORR").
ENV["TYPE"] = "CORR"
# Sets the dimension of the problem.
ENV["DIMENSION"] = "100"
include("run_optimal_design.jl")
For individual instances, you can call the solvers directly.
using ODWB
seed = 1
m = 100
n = 10
criterion = "D"
time_limit = 3600 # 1 hour
corr = true # this create correlated data, if false, it will create independent data.
# Boscia
ODWB.solve_opt(seed, m, n, time_limit, criterion,corr)
# SCIP
if criterion in ["A", "D"]
error("SCIP OA does not work with the optimal problems!")
end
ODWB.solve_opt_scip(seed, m, n, time_limit, criterion, corr)
# Pajarito
ODWB.solve_opt_pajarito(seed, m, n, time_limit, criterion, corr)
# Custom
ODWB.solve_opt_custom(seed, m, n, time_limit, criterion, corr)