Eine logische Applikationssprache zur parallelen Verarbeitung von Neuronalen Netzen - Anwendung in der Qualitätssicherung von Agrarprodukten

This thesis is a contribution to the fields of distributed artificial intelligence, parallel logical programming and parallel processing of neuronal networks on a cluster of workstations. A logical application language is described which was designed specifically for quality assurance for agricultural products. The application language supports access to databases. The logical application language Mulog consists of several components which communicate via CORBA and SOAP. Moreover, the generating of neuronal networks on several computers in a computer network is supported, and parallel execution of programs is allowed. The source code of the programs written in the application language Mulog is translated by a compiler into an intermediate code which is executed in program run-time by an intermediate code interpreter. Parallelisation is achieved by an extension of the intermediate code commands which permits asynchronous process control and distribution of the intermediate code programs to a cluster of computers in run-time. The development of knowledge-based systems is supported by the inference machine. A special form of the first-order predicate calculus is the presentation as a Horn clause. Horn clauses forms the basisi of the syntax of the logical application language Mulog. The whole system allows the dynamic addition of computers in run-time. Different types of neuronal networks can be generated directly by commands in the application language Mulog. Black spots arise in potatoes during harvesting, storage and processing. Many consumers are familiar with these black spots, which are regarded as a quality defect. To enable this quality defect to be remedied, the influencing factors that cause it must be identified. A database with potential suspected influencing factors such as mechanical stress, temperature, potassium etc. was collected for different potato varieties. With the aid of the application language Mulog, programs were developed which generated a large number of neuronal networks parallel to each other on several computers in a computer network. A large number of back-propagation neuronal networks were created for each potato type. The input vector consists of the data for two influencing factors. The output vector is the measured incidence of black spots. All permutations of two influencing factors are used as the basis for each neuronal network, so that at the end there is a neuronal network for each variety of potato for each possible combination of two influencing factors. These neuronal networks were then tested. By comparing the black spot prediction by the neuronal networks with the actual incidence of black spots, conclusions can be drawn about the influencing factors used to generate the data. The evaluation of the data developed and executed by the programs in the Mulog system showed an agreement with the current state of knowledge on the influencing factors. In addition, influencing factors which are also regarded as causes of the black spots could be evaluated in their importance for the origination of black spots. An evaluation of individual potato varieties shows the special characteristics of the individual varieties.