Comparative Testing

• Methods:

  • Simulated Annealing: AMEBSA, ASA, SALO
  • Multi Level Single Linkage (MLSL) and variants
  • Random Local Optimization (RANDLO)

• Test Functions:

  • From optimization literature and method demos
  • Used to gain rough idea of relative strengths

Previous slide

Next slide

Back to first slide

View graphic version

For this study, we've chosen to compare three main types of global optimization methods: simulated annealing, multi level single linkage (MLSL), and random local optimization. AMEBSA is simulated annealing based on the downhill simplex method. ASA is Lester Ingber's adaptive simulated annealing which uses importance sampling to adapt annealing schedules separately for each parameter. SALO is a simulated annealing method (ASA in this case) applied to a function space transformed by local optimization. Multi Level Single Linkage (MLSL) performs local optimizations from a subset of uniformly sampled points. Variants of MLSL make use of lazy function evaluation and assumptions of determinism in local optimization procedures. RANDLO is simply random local optimization.

Test functions were chosen from global optimization literature and algorithm demonstrations in order to roughly assess the relative merits of these global optimization methods. The paper contains details about the methods, the local optimization procedures they use, and the test functions.