New Meta Algorithms for Engineering Design Using Surrogate Functions
Abstract
We have recently made an important breakthrough in both the theory and the practice of derivative-free methods. Our new research eliminates the main limitation of the Generalized Pattern Search (GPS) class of algorithms, which formed the algorithmic infrastructure for our past work. Our new work addressed the issue that the OPS algorithm class relies heavily on a fixed finite set of directions, and that the convergence results are heavily tied to these directions. We found a way to lift these restrictions, and now we allow the algorithm to explore in a set of directions which is dense in the entire space. We call this the Mesh Adaptive Direct Search (MADS) class of algorithm, and it generalizes GPS. we observed significant improvements on some interesting problems.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 2005
- Accession Number
- ADA433179
Entities
People
- John E. Dennis Jr.
Organizations
- Rice University