Dominik Wieland

FROM NEURAL NETWORKS TO QUALITATIVE MODELS IN ENVIRONMENTAL ENGINEERING

As an alternative to physical models, artificial neural networks (ANNs) are a valuable forecast tool in environmental sciences. They can be used effectively due to their learning capabilities and their low computational costs. Once all relevant variables of the system are identified and put into the network, it works quickly and accurately. However, one of the major shortcomings of neural networks is that they do not reveal causal relationships between major system components and thus are unable to improve the explicit knowledge of the user. Another problem is due to the fact that reasoning is only done from the inputs to the outputs. In cases where the opposite is requested (i.e., deriving inputs leading to a given output), neural networks can hardly be used. To overcome these problems, we introduce a novel approach for deriving qualitative information out of neural networks. Some of the resulting rules can directly be used by a qualitative simulator for producing possible future scenarios. Because of the explicit representation of knowledge, the rules should be easier to understand and can be used as a starting point for creating models wherever a physical model is not available. Moreover, the resulting rules are well adapted to be used in decision support systems. We illustrate our approach by introducing a network for predicting surface ozone concentrations and show how rules can be derived from the network and how the approach can be naturally extended for use in decision support systems.

Comparing Two Models for Software Debugging

This paper extends previous work on the representation and analysis of Java programs for diagnosis in a new direction by providing a description and analysis of the issues arising from handling object references in dependency-based models of Java programs. We empirically compare dependency-based models with a value-based model using a set of example programs in terms of required user interaction (questions put to the user) and examine and incorporate specific interesting error categories. Apart from being based on experience with an actual implementation of the various models, the model extensions and analysis deal with aliasing, an issue that the programming language community has been examining for a long time, and that is also crucial to object-orientedness.