• Profile
• Contact us
• Methods
• Projects
• Publications
• Links

The structure as well as the functionality of artificial neural networks (ANNs) are inspired by biological nervous systems: many simple units called artificial neurons, which can process simple nonlinear tasks only when they are isolated, are combined to a complex network which is capable of approximating complex nonlinear functions.

According to learning mechanisms in nervous systems, an ANN also has to learn a task on the basis of example data. This is called the training phase or learning phase of an ANN. Different learning algorithms are applied which can be classified into supervised and unsupervised ones. The example data an ANN is given during training has to be preprocessed.

After the training phase, an ANN will deliver approximated solutions for nonlinear and stochastic correlations or it will solve classification and controlling tasks.

Using ANNs makes it possible to conquer nonlinear functions without the need of setting up any model of the underlying real world process which is often very time consuming.

• Learning ability (supervised and unsupervised learning).
• Generalization.
• Approximation of nonlinear dependencies.
• Classification ability.
• System design without a priori knowledge.
• Visualization of data structures.

By being trained with example data, ANNs learn to approximate a certain non-linear function.

As this example data does not only contain the information relevant for the function but also various other (irrelevant) signals, certain data preprocessing steps have to be carried out before the actual training starts.

During those steps the relevant parts of the data are identified, isolated from the irrelevant signals and amplified. Frequently, filtering algorithms are applied. This usage of data preprocessing methods allows an estimate for the quality of the example data.

In many cases, it is sufficient for the solution of a nonlinear problem to consider only a few parameters - even when a variety of parameters could possibly be used.
By applying data preprocessing algorithms, those parameters who contribute best to a solution are identified.
This reduction of free parameters leads to an efficient reduction of data space, which results in faster process time and more precise solutions.