next up previous contents
Next: 4.5 Integration into a Up: 4. Optimizer Previous: 4.3.6 Model Library


4.4 General Nonlinear Optimization Problems

General nonlinear optimization problems have a slightly different structure than the least-squares problems discussed in the previous section. Here the cost function $f(\vec{x})$ should be minimized. For doing this the optimizer can vary the parameter vector $\vec{p}$ within the defined boundaries.

In this general description of the problem it can occur that a numerically optimal solution lies within a nonphysical area. For this reason the feasible area has to be reduced. This is done by so called inequality and equality constraints.

Figure 4.4: Structure of a general nonlinear optimization problem with equality and inequality constraints.
\resizebox{6.cm}{!}{
\psfrag{Model}{Model}
\psfrag{M}{$\mathcal{M}$}
\psfrag{...
...f}{Target Function $f$}
\includegraphics[width=6.cm]{graphics/gnostruct.eps}
}

The solution of a general nonlinear optimization problem is found using the package donlp24.5 in which the following steps are processed.

1. Choose a value $\delta_m$

2. From all inequality constraints, only those are selected which are nearly active ( $g_i(\vec{x}) \le \delta_m$).

3. Solve the quadratic problem built by the gradient and Hessian of the target function subject to a reduced set of inequality constraints and all equality constraints. The Hessian was built by a BFGS-update described in Section 4.1.6. The result is the step direction and the slack variables u and v.

4. Check for termination

5. If the calculated step is to small or the violation of the constraints exceed tau0 then increase $\tau_m$ and go to step 3 else proceed.

6. Compute new penalty weights

7. Select the step size

8. Assign a new value for x and $\tau$ and resume at step 1



Footnotes

...donlp24.5
Further details of the algorithm can be found in [49,50] or in the documentation of the program in [51].

next up previous contents
Next: 4.5 Integration into a Up: 4. Optimizer Previous: 4.3.6 Model Library

R. Plasun