International Symposium
on Mathematical Programming 2000
Atlanta, GA
August 7-11, 2000
Pattern search finds the minimum of a function using only function values and is amenable to parallel computing because it requires multiple function evaluations, which can be performed simultaneously, at each iteration. The problem is that all function evaluations from one iteration must complete before the next iteration can proceed. On many types of parallel computers and for many types of objective functions, the computation times vary greatly from processor to processor leaving many processors idle. An asynchronous pattern search method eliminates this problem by not requiring any synchronization. We will describe this new method, discuss related convergence theory, and demonstrate its effectiveness of several applications. This is joint work with Virginia Torczon, College of William & Mary.
SAND2000-8833A.
postscript or pdf
Maintained by: Tamara Kolda
(tgkolda@sandia.gov).
Disclaimer and Acknowledgment.