Projected-Search Methods for Box-Constrained Optimization
Michael Ferry
UCSD
Abstract:
We survey several commonly-used quasi-Newton methods and line-search algorithms
for unconstrained and box-constrained optimization and consider their underlying
strategies. By taking advantage of an implicit similarity in two existing algorithms, we develop a method for box-constrained optimization that includes a
new way to compute a search direction and a new line-search algorithm. On a
collection of standardized problems, this method is over 35% faster than the
leading comparable alternative.