EUFIT '99

Tutorials

Applications, September 13, 1999, 16:30 - 18:30 h

I: Process Industry

Timo Slawinski, University of Dortmund

Fuzzy technologies and neural networks are increasingly being used in process industries. Due to the complexity and non-linearity of most of these processes, exact mathematical models are often difficult or even impossible to establish. In the tutorial, the different methods will be compared exemplary due to their ability to deal with real world applications. Typically, robust methods are required, which are able to manage contradictory data and complex high dimensional search spaces. Additionally, often transparent and comprehensible results are desired to reach a better acceptance in the process industry.

If there is process knowledge available, fuzzy systems enable experts to exploit their qualitative knowledge by using "if-then-rules" (knowledge based approach). It will be demonstrated, that positive rules can be interpreted as recommendations and negative rules can be interpreted as warnings and prohibitions to avoid undesirable operation situations. As direct design by experts is often not possible, the application of fuzzy models and fuzzy controllers depends on efficient data-based methods. Different approaches and applications will be discussed in the tutorial. As an alternative approach in the field of data-based modelling neural networks will also be considered. Examples for the successful application of neural networks for different type of real world tasks are presented and compared with fuzzy technologies.

The following applications will be presented in the tutorial:

• an alkoxylation process (example for modelling human behaviour)
• adaptive controller design in robotics (example for learning from good and poor control strategies)
• classification of automatic gear boxes (example for rule generation in high dimensional search spaces)
• load prediction (example for learning from very noisy and contradictory data)
• prediction of cement quality (decision support system for a plant operator)

J: Finance and Trade

Hans-Georg Zimmermann, Siemens AG, Germany

This workshop presents several methods to incorporate econometric and financial knowledge into a neural network modeling environment. Therefore, the outcome of the model building does not only depend on the data but also on the incorporated economic structure.

In a sequence of steps we will discuss topics in forecasting financial indices, strategic portfolio management, yield curve forecasting and stock picking.

In a first study we will analyze 6 month forecasts of stock indices and bonds by feedforward neural networks. A simple formulation of this task is the description of the present state of a market by a state vector and the attempt to relate a future state of the market as an output. But this naive pattern recognition approach is totally neglecting the fact that we try to analyze a dynamical system. We have to ask the question how we can include the dynamical aspect in our network architecture. In the first section we will work out an answer to this question. A note on learning techniques will finish this section.

In the next section we will take the above experiences as a basis of a strategic portfolio management system. Here we have not only to develop forecast models for many assets in 7 counties but we will see how the transformation from the forecast into an asset allocation schema can be done. This approach will be compared to the classical Markovitz and the Black- Litterman approach. We will end up not only with a neural network which combines forecasting and decision support but we also get a new view on the economics/econometrics of the problem itself.

The next section is focused on yield curve forecasting. Instead of modeling every maturity in an individual model we will analyze a neural network architecture which allows the modeling of many outputs if they belong to a common underlying dynamic.

In a following section we will combine the the ideas from the portfolio optimization and the yield curve analysis to a tactical portfolio management system or to say it in other words to a stock picking decision support system. In contrast to the strategic portfolio problem where we have to analyze the asset allocation for a relative small number of assets (stock indices or bonds in different countries) driven by very different dynamics all around the world, we now have the difficulty to work with a large number of assets (e.g. the stocks included in one of the stock indices in one country) assuming that their behavior is driven by a combined underlying dynamic.

In a last section we will discuss some future developments. From the econometric view the use of recurrent neural networks will give us not only a new view on the analysis and modeling of dynamical systems but it provides the right frame for new questions, e.g. the search for early warning indicators. From the economic view it is interesting to study neural network based multi agent systems. They allow an explicit modeling of the price mechanism.

The aim of the workshop is not only to give you an overview on selected neural network techniques but also to provide an understanding how neural network projects in economics should be organized and what are the critical points for the success of such projects. Siemens has developed in the last decade a software environment called SENN which allows a flexible design of all the mentioned tasks.

K: Medicine and Healthcare

Derek Linkens, University of Sheffield, UK

The human body is a hugely complex, non-linear dynamic organism, whose functions are only partially understood. Unlike the realm of physical sciences where the basic laws can be applied to obtain accurate quantitative models, in the life sciences our knowledge is often restricted to qualitative and heuristic relationships. In this field, the computational techniques of fuzzy logic, neural networks and genetic algorithms offer significant potential, both for scientific understanding and engineering application in medicine and healthcare.

In this Tutorial, an overview will be given of the penetration of soft computing into different sub-regions of medicine in recent years. This will be followed by an in-depth case study of the use of the three main computational intelligence paradigms in the field of anaesthesia in operating theatres. The sub-areas of unconsciousness and muscle relaxation will be used to illustrate the use of self-organising fuzzy control, neuro-fuzzy patient modelling, neural network based monitoring, and genetic algorithm optimisation. Methods of hierarchical decomposition and data-fusion will be described for these applications where both definitions and measurements are difficult and often imprecise.

- EUFIT '99 -