Parallel Jacobi methods for derivative-free optimization on parallel or distributed processors

Abstract

New Jacobi-type algorithms are presented for the efficient use of parallel and distributed computing platforms in solving derivative-free optimization problems. The implementations are designed to be fault tolerant to be applicable to science and engineering problems where occasionally requests for function values may not be met and derivatives are never available. Convergence is usually achieved by introducing an elementary trust region subproblem at synchronization steps in the algorithm. This has the added advantage of handling negative curvature very conveniently.