|
KNITRO - Optimization techniques
The optimization techniques used by KNITRO offer the leading
combination of computational efficiency and robustness.
KNITRO implements three different approaches,
advantageous on models with different structures:
•
handling of inequality constraints by an interior point algorithm
and direct solution of the barrier subproblems. This strategy
is especially recommended for ill-conditioned problems;
•
handling of inequality constraints by an interior point algorithm
and solution of the barrier subproblems by conjugate gradient
iterations. This approach is recommended for large-scale problems
with dense Hessians;
•
handling of inequality constraints by an active set algorithm,
which is especially beneficial when a "good" initial
point is available and when solving a sequence of related
problems. This option is also recommended for detecting infeasibility.
By default, KNITRO will try to make an automatic selection
of the best algorithm to use based on the problem characteristics.
These three approaches derive from Newton's method. KNITRO's
overall global convergence properties are ensured by the use
of trust regions.
Other features
The evaluation of second derivatives is optional. KNITRO offers
quasi-Newton (BFGS) and finite difference options that approximate
the second derivatives of the model and therefore relieve
the developer of their computation.
A specialized function of KNITRO makes it possible to take
the particular structure of linear and quadratic models into
account and further improve performance. Mathematical problems
with equilibrium constraints, which arise for instance in
energy markets, can be explicitly handled by KNITRO. This
specialized version of KNITRO algorithms enable an efficient
solving of MPEC.
A new crossover feature has been added. When a problem has
been solved the interior-point algorithm, the crossover procedure,
which is an active set iteration, identifies the active set
at the optimum and gives accurate estimates of the Lagrange
multipliers.
A "feasible" mode is available. It ensures that,
as soon as KNITRO has identified an iterate satisfying the
inequality constraints - in some cases, the initial guess
provided by the user - then all subsequent iterates also satisfy
these constraints. This is particularly useful when the model's
functions are not defined everywhere.
KNITRO runs on Unix (Sun Solaris), Linux and Windows platforms.
Contact
For more information regarding KNITRO,
please contact us at:
•
Tel: +33 1 44 77 89 00
•
E-mail: info-knitro@artelys.com
Back to
the top
|
