SOLNP Test Suite

Introduction

The purpose of the new SOLNP Test Suite is to provide a comprehensive, standardized environment for benchmarking and validating nonlinear optimization algorithms, with a particular focus on the SOLNP family of solvers. The suite is anchored around the well-known Hock-Schittkowski (HS) test problems, a diverse collection of nonlinear programming challenges that have become a benchmark standard in the optimization literature.

This suite currently implements 64 of the HS problems, spanning a range of complexities, constraint types, and objective function characteristics. Future updates will expand coverage to include the entire set of 306 Hock-Schittkowski problems, ensuring a thorough and robust testing ground for both algorithmic development and comparative analysis.

The problems are currently stored in a form compatible with the SOLNP solver, enabling immediate use for benchmarking and analysis. However, recognizing that nonlinear programming (NLP) problems can be specified in various formats—sometimes with differing conventions for constraints, bounds, and objective function representation—the suite includes a utility function called solnp_standardize_problem which rewrites the test in the NLP standard form (e.g. as used by nloptr).

A distinctive feature of several HS problems is that their optimal solutions often occur on or very near the boundary of the feasible region, where one or more inequality constraints are active. In other words, the optimal solution frequently coincides with points where the constraints are “tight”, and further movement would violate feasibility.

This aspect is of particular importance for algorithm development, as it tests a solver’s ability to:

Correctly identify the active set of constraints at the optimum,
Handle boundary solutions robustly, and
Maintain numerical stability and feasibility in the presence of active inequality constraints.

It should be noted that SOLNP does not employ an active set strategy. Instead, its augmented Lagrangian approach treats constraints via penalty terms, and the optimizer does not explicitly identify or maintain sets of active constraints during the iterations. As a result, SOLNP solutions typically remain in the interior of the feasible region whenever possible, only approaching the boundary when required for optimality.

Usage

The function solnp_problems_table lists the problem currently implemented:

Code

head(solnp_problems_table())
#>               Suite Problem Number
#> 1 Hock-Schittkowski    hs01      1
#> 2 Hock-Schittkowski    hs02      2
#> 3 Hock-Schittkowski    hs03      3
#> 4 Hock-Schittkowski    hs04      4
#> 5 Hock-Schittkowski    hs05      5
#> 6 Hock-Schittkowski    hs06      6

whilst the function solnp_problem_suite can either return a specific problem or all problems within a Suite:

Code

prob <- solnp_problem_suite(suite = "Hock-Schittkowski", number = 60)
sol <- csolnp(prob$start, fn = prob$fn, gr = prob$gr, eq_fn = prob$eq_fn, 
              eq_b = prob$eq_b, eq_jac = prob$eq_jac, ineq_fn = prob$ineq_fn, 
              ineq_lower = prob$ineq_lower, ineq_upper = prob$ineq_upper, 
              ineq_jac = prob$ineq_jac, lower = prob$lower, upper = prob$upper, 
              control = list(trace = 0, min_iter = 1000, max_iter = 300, tol = 1e-8, rho = 1))
print(c("solnp" = sol$objective, "benchmark" = prob$best_fn))
#>     solnp benchmark 
#> 0.0325682 0.0325682
print(rbind("solnp" = sol$pars, "benchmark" = prob$best_par))
#>               [,1]     [,2]     [,3]
#> solnp     1.104859 1.196674 1.535262
#> benchmark 1.104859 1.196674 1.535262