BTN - Block truncated Newton

Unconstrained nonlinear minimization for parallel computers. Suitable for large-scale optimization.

BTN uses a block truncated Newton method in conjunction with a line search strategy. An approximate Newton direction is obtained by applying the block conjugate gradient method to the Newton equations. Blocking is used to enable parallelism in both the linear algebra and the function evaluations.

Both easy-to-use and customized versions are provided. The easy-to-use version requires only that the user provide a (non-parallel) subroutine to evaluate the objective function and its first derivatives; no knowledge of parallel computing is required. The customized version allows more complicated usage, including parallel function evaluation.

A parallel derivative checker is also included. The software can be run on traditional computers to simulate a parallel computing environment.

The software is written in ANSI Fortran 77, using double precision real variables. A small number of machine-dependent subroutine calls and compiler directives control the parallel execution. Machine constants are set in a single subroutine (d1mach). Versions of the software are available for the Intel iPSC/2 and iPSC/860 hypercube computers (distributed memory) and the Sequent Balance and Symmetry parallel computers (shared memory). The Sequent version can also be run on serial computers to simulate the performance of a parallel machine.

The code is available from from Netlib.

Need more info?

Contact:

Stephen Nash
ORAS Department
George Mason University
Fairfax VA 22030.
Phone: (703) 993-1678.
snash@gmu.edu

References:

S. G. Nash and A. Sofer, BTN: Software for parallel unconstrained optimization, ACM Trans. Math. Software 18 (1992), pp. 414--448.

S. G. Nash and A. Sofer, A general-purpose parallel algorithm for unconstrained optimization, SIAM J. Optim. 1 (1991), pp. 530--547.

[ Optimization Software Guide | OTC Home Page | NEOS Server | NEOS Guide ]