Neural Network Projects

The bayesian framework provides a probabililistical model to improve the generalization performance of neural networks. Starting from the Bayes'schen formula, we derive an a-posteriori probability distribution, which in addition to the training data considers a-priori information about the learning and optimization process. Below acceptance of normal distributed weights, this technique leads us to a well-known error function, which is learning with weight-decay. We are occupied with the improvement and enlargement of the existing algorithms. Main focus is a suitable modeling of the a-priori probabilities of the weights and the hyperparameters. In several projects, e.g. forecasting financial time series, the bayesian learning procedure could be used successfully.

(S. Gutjahr)  

The goal of neural network training ist not to represent the training data exactly, but to generate a statistical model of the underlying (data generating) process. When designing neural systems several methods are taken into account to guarantee the generalization capability of the model, e.g., regularisation (cf. Bayesian learning), feature extraction and topology optimization as well as committees of networks.

The program ENZO integrates several of these paradigms into an optimization algorithm. In analogy to the example of natural evolution, a selection pressure is used to create high performant models (e.g. with high evidence). The DFG research project "Integrierte Entwicklung von Komitees neuronaler Netze" (Integrated design of committees of neural networks) has the goal to design with this evolutionary optimization algorithm different neural models which are as independent as possible in the mathematical sense. These neural models can be combined to a committee in a sensible way to increase the generalization performance of the whole system.

(T. Ragg)  

Forecasting of economic time series prepares considerable difficulties because the compatibility of cause and effect is highly complex and the data is very noisy. Therefore it is not possible to detect the complicated market mechanisms by means of linear approaches. In recent years neural networks were used more and more, which are more suitable to make forecasting on account of their non-linear structure.

In cooperation with the state bank Hessen-Th"uringen, we have developed prediction systems for the daily forecast of financial time series since 1994. As input we use, so-called technical indicators, i.e. characteristics that are computed from the historical price movement of the underlying time series. On the basis of the forecast of the neural network, a commercial system is developed which maximizes the profit. This procedure was applied to different future and stock prices as well as to the dollar DM exchange rate and is used successfully in the bank for daily trading.

In August 1998, a project with the Axel Springer publisher was started in order to predict the daily sell-rates for the newspaper 'Bild'. The aim is to minimize the reflux of the newspapers, in the case of identical sale numbers.

(S. Gutjahr and T. Ragg)  

The DFG-project FYNESSE aims at developing an autonomously learning hybrid control architecture for unknown dynamic processes. The learning problem is characterized by the absence of any a priori knowledge about the system behaviour. The training information that is available for the controller merely consist in the specification of the control goal.

The concept is based on the cooperation and interaction of two adaptive modules, namely a neural critic module and a fuzzy control module. While the control strategy is realised in the interpretable fuzzy component the purpose of the neural net lies in evaluating the consequences of the current strategy. This enables an incremental improvement.

(M. Riedmiller and R. Schoknecht)  

The interdisciplinary research project "Information Structures in Music", which currently is supported by the Klaus Tschira-Foundation, applies methods used in computer science to analyze and model musical structures. One research interest concentrates on rule-based techniques (e.g. formal grammars), another uses learn-based approaches, especially artificial neural networks and genetic algorithms.

Modeling musical structures with rule-based procedures is sufficient for the description of general composition principles, but is not suitable to characterize historical or individual composition styles. By their adaptive behavior, neural networks however are able to learn style-typical features directly from musical examples and can imitate the intuitive capturing of musical structures. Thus they are able to produce convincing solutions also in unexpected situations, e.g. to harmonize new melodies and provide them with melodic variations in the style of a certain composer.

(D. Hoernel and K. Hoethker)  

Robot Soccer is a benchmark for the application of machine learning techniques in large complex multi-agent environments. Our research focuses on the application of reinforcement learning: As training information, only the final objective is provided (shoot one goal more than the opponent). We try to develop distributed learning algorithms, that eventually are able to learn a cooperative and succesful team strategy.

For further information also visit our homepage.

(M. Riedmiller)