QML-Morven : a framework for learning qualitative models

The work proposed in this thesis continues the research into qualitative model learning (QML), a branch of qualitative reasoning.  After the investigation of all existing qualitative model learning systems, especially the state-of-the-art system ILP-QSI, a novel system named QML-Morven is presented. QML-Morven inherits many essential features of the existing QML systems: it can learn models from positive only data, make use of the well-posed model constraints, process hidden variables, learn models from incomplete data, and perform systematic experiments to verify the hypotheses being made by researchers. The development of QML-Morven allows us to further investigate some interesting yet unsolved questions in the QML research.  As a result, four significant hypotheses are tested and validated by performing a series of systematic experiments with QML-Morven:  1. The information of state variables and the number of hidden variables are two important actors that can influence the learning, and the different combination of these two factors may give a different learning result in terms of the kernel subset (minimal data for a successful learning) and learning precision; 2. The scalability of QML may be improved by the use of an evolutionary algorithm; 3. For some models, the kernel subsets can be constructed by combining several sets of qualitative states, and the states in a kernel subset tend to scatter over the solution space; 4. The integration of domain-specific knowledge makes QML more applicable for learning the qualitative models of the real-world dynamic systems of high complexity. The results and analysis of these experiments with respect to QML-Morven also raise many questions and indicates several new research directions.  In the final part of this thesis, several possible future directions are explored.EThOS - Electronic Theses Online ServiceGBUnited Kingdo