Next: 6.3 Stability
Up: 6. Discussion
Previous: 6.1 Optimization in a
In order to setup an analysis task the following steps have to be done:
- Formulation of the goal of the analysis.
This results in a set of parameters
which probably have an effect on one ore more output parameters of a
simulation model.
- Assembly of the model function.
The analyzed model must be assembled
by using several simulation or data manipulations tools. Special
care must be taken on numerical noise of the simulators. This noise
can be caused especially from regridding steps, Monte-Carlo
simulations etc.
Also truncation errors in the results can introduce undesirable
nonlinearity into the optimization loop.
- Performing the analysis.
In this step the framework does the majority
of the job. The user only has to keep an eye on the availability of
computational resources such as disk space,
should frequently check the parameters
and the simulated data. Thereby errors contained in the setup of the
optimization can be identified early.
- Discussion of the results.
The results of the optimization have to be
verified carefully.
When the values correspond with the expected results it is an
indication for a successfully analysis.
Checking of the data generated during the target function
evaluations of critical input parameters is necessary to assure the
correctness of the results.
- Adapt the model or the ranges of the parameters.
If the results are not satisfying the source can be usually found in
the model or input parameters.
Typical examples for systematic mistakes are caused by choosing sets of input parameters
not suitable for the type of analysis, e.g., missing constraints,
which pushes the optimization away from unphysical results.
Setting up an analysis task can take some time but this time is
invested very well. The user gets insight in the dependencies of the
inputs and responses and is able to obtain optimum values with a
minimum of user interaction.
Next: 6.3 Stability
Up: 6. Discussion
Previous: 6.1 Optimization in a
R. Plasun