Search

Direct search optimization algorithms are becoming an important alternative to well-established gradient based methods. Due to the fact that a single cost function evaluation may take a substantial amount of time, optimization can be a lengthy process. In order to shorten the run time one often resorts to parallel algorithms. Asynchronous algorithms are particularly efficient since they have no synchronization points. This paper is an attempt to establish a convergence theory for a class of such parallel direct search algorithms. The notion of a search direction generator (SDG) is introduced. An algorithmic framework for parallel distributed optimization methods based on SDGs is presented along with the corresponding convergence theory. The theory almost completely decouples the stepsize control from the sufficient descent requirement, which is necessary for the finite termination of the algorithm's inner loop. The proposed framework has several attributes considered very favourable in loosely coupled parallel systems (e.g. clusters of workstations), such as fault tolerance and scalability. The framework is illustrated by optimizing a set of test problems on a cluster of workstations. In all tested cases, a speedup was obtained that increased with the increasing number of workstations. Fault tolerance and scalability of the framework were also demonstrated by removing and adding workstations to the cluster while an optimization run was in progress. [ABSTRACT FROM AUTHOR]