1999-03-09
16:15 at IFW A32

An Asynchronous and Fault-Tolerant Algorithm for Parallel Direct Search Optimization

Tamara G. Kolda

Oak Ridge National Laboratory

We develop a parallel direct search algorithm for unconstrained optimization that is completely asynchronous (or chaotic) as well as fault-tolerant. We are motivated by problems that have only a few variables (say 10-50) but very expensive objective functions (such as a complex simulation). Our target parallel architecture is a cluster of workstations, possibly heterogenous, with varying loads and reliability --- the only `supercomputer' available to a typical scientist. So, we desire an unconstrained optimization method that requires no derivatives, can be parallelized for a heterogeneous environment, handles node failures, and still guarantees global convergence