Performance#

Performance profile and data profile#

Unconstrained optimization#

This section compares smt-optim (EGO) with COBYLA on five global optimization problems. For each problem, five instances are proposed, each with a unique design of experiment (DoE), for a total of 25 instances. The data profiles below are used to compare the required budget to $\tau$-solve an instance. Note that COBYLA is a derivative-free local optimizer, whereas EGO is a global optimizer. The purpose of the benchmark below is to demonstrate the global convergence of EGO.

The five global optimization problems are the following:

  • Ackley

  • Griewank

  • Levy

  • Rastrigin

  • Schwefel

The data profile is for the two-dimensional representation of the five selected problems. Each initial DoE comprises three points. Cobyla started from the point with the lowest objective value.

image

Constrained optimization#

Multi-fidelity optimization#