The main motivation for developing the software was to evaluate practically
the suggested model. For bigger neural networks it is very difficult to
analyze the behaviour of the system theoretically, in particular the
generalization performance. One way to demonstrate that the model is working
is to run a simulation. This does not formally prove that
the proposed architecture has certain properties because
only a finite number of examples can be simulated. Nevertheless this methodology
is widely used in the field of neural networks. From a more practical point
of view a successful simulation does *`prove'* that the architecture
is capable of solving a certain type of problem.

In this section aspects of the implementation are discussed, including the structure of the software for the modular neural network, the usage of the programs and the interfaces to the developed library. The implementation also included mechanisms for preprocessing of the data.

- The Concept
- The Software Platform
- Preprocessing of the Data
- The Neural Network Library
- The File Formats
- The Developed Programs

Mit Okt 4 16:45:34 CEST 2000