Dynamic adaptive search for large-scale global optimization

Conference · Sat Dec 31 04:00:00 EST 1994

OSTI ID:36146

Hu, T C; Kahng, A B

We directly exploit a new, more realistic paradigm for the iterative optimization process itself, wherein we return the best-ever solution found over the entire computation rather than the last-seen solution that is generated in the final iteration. We propose non-monotone, adaptive threshold methods which are self-tuning to the individual optimization instance. These methods allow efficient escape from local minimum solutions because the parameter (threshold or temperature) schedule is allowed to be highly non-monotone, in some sense {open_quotes}following{close_quotes} the algorithm`s knowledge of the cost surface. We propose strategies for bounded-time optimization, i.e., strategies which explicitly depend on the total computing time allowed as well as the current stage of the optimization. We exploit a new {open_quotes}big valley{close_quotes} picture of structure in the optimization cost surface to generate effective initial states for the computation; these ideas greatly influence the implementation of {open_quotes}multi-start{close_quotes} strategies.

• OSTI ID: 36146
• Report Number(s): CONF-9408161--
• Country of Publication: United States
• Language: English

Related Subjects

99 GENERAL AND MISCELLANEOUS
ALGORITHMS
ITERATIVE METHODS
LINEAR PROGRAMMING
MATRICES
NUMERICAL SOLUTION