Comparative Study Results (cont.)
For test functions STEP1 and STEP2, RANDLO and LMLSL performed best for both constrained local optimization procedures.
SALO: ASA did not search the locally optimized search spaces (f´) efficiently.
Recent experiments indicate that information-based global optimization performs even better.
Notes:
We call the test function shown STEP1 and use a logarithmically scaled version of it, STEP2, for testing as well. Two different constrained local optimization procedures, CONSTR and YURETMIN, were used for variety.
Results indicated that RANDLO and LMLSL (MLSL with lazy function evaluation) performed best for STEP1 and STEP2 respectively using each of the local optimization procedures.
It’s interesting that SALO didn’t do as well for these functions. Most of our optimization methods, including SALO, have global and local phases. Local optimization of function f is often performed as a subroutine so that the global phase is effectively searching f’, a simpler landscape of plateaus. The local nature of simulated annealing on f’ proved problematic in these cases.
It’s also interesting to note that most of the information gained in the course of optimization is thrown away. RANDLO throws away all information except the lowest minima it has found. The method which retains the most information, MLSL, keeps only that and the sampled function evaluations it collects in its global phase. All methods throw away all but the final function evaluation of local optimization. More recent results indicate that information-based optimization methods, which use such information optimally, fare better with test functions such as these.