Computer-Supported Experiment Selection for Model Discrimination

No part of this work may be reproduced by print, photocopy or any other means without the permission in writing from the publisher. Before him [Gandalf] stood a dark arch opening into three passages: all led in the same general direction, eastwards; but the left-hand passage plunged down, while the right-hand climbed up, and the middle way seemed to run on, smooth and level but very narrow. Summary Obtaining an adequate model of an experimental system is a laborious and error-prone task. Model construction involves the analysis of observations of system behavior, as well as the application of domain knowledge. In this process, different assumptions about the structure and behavior of the system can be made. Often, this results in a number of competing models, each equally justifiable by the available observations. In order to discriminate between the competing models, new observations on the system behavior have to be made. These can be obtained by performing additional experiments on the system. Since in real-life applications the number of experiments that can be performed may be quite large, and the cost of each of them considerable, it is important that the experiments be selected carefully. It is preferable to select experiments in such a way that the set of possible models is maximally reduced at minimal costs. The selection of experiments for model discrimination has received attention in the statistical literature. The methods described there apply to mathematical models with a precisely-defined structure and exact numerical values for the system variables. However, in many situations, incomplete or imprecise information about the system does not allow the formulation of quantitative models. This fact has motivated the work described in this thesis, the development of a method for the selection of experiments to discriminate between semi-quantitative models. In a semi-quantitative model imprecise and incomplete information is represented by numerical intervals bounding parameter values, and by envelopes bounding unknown functions. Experiment selection is based on the predictions of the competing models that have been obtained through simulation. For the formalization and simulation of semi-quantitative models, techniques developed in the field of Qualitative Reasoning have been employed. More specifically, the technique QSIM and its extensions Q2 and Q3 have been used. Simulation with QSIM, Q2, and Q3 infers from a model a set of possible semi-quantitative behaviors of an experimental system. A semi-quantitative behavior is a qualitative behavior augmented with interval information for the values …