Title: Hybrid Optimization Parallel Search Package Speaker: Josh Griffin, Sandia National Laboratories Date/Time: Wednesday, September 19, 2007, 10:00 am - 11:00 am Location: CSRI Building, Room 90 (Sandia NM) Brief Abstract: The focus of this talk will be on a new software platform being developed that facilitates the formation of tightly coupled hybrid optimization algorithms using native source code. Currently two levels of parallelism are supported where both function evaluations and individual optimization routines may run asynchronously in parallel sharing a single function value cache and prioritized evaluation queue. Because each solver comes with inherent strengths and weaknesses, it is difficult to determine beforehand which solver is best for a given problem or region. Further, initially little may be known of the characteristics of the problem itself. By combining algorithms in parallel we can exploit load balance to create efficient hybrids that share information, helping each other to hop past rather than converge to individual trouble spots. Ultimately the goal of this software is to offer a platform where the user can easily and dynamically building robust parallel hybrid solvers, moving the burden and guesswork of choosing the appropriate routine from the user to the software package itself. We provide preliminary numerical results for LHS/GSS/DIRECT hybrids for global optimization, and hybrid exact penalty methods for nonlinearly constrained optimization.CSRI POC: Brian Adams, (505) 284-8845 |