Search

This paper addresses the issue how to strike a good balance between accuracy and compactness in classification systems - still an important question in machine learning and data mining. The fuzzy rule-based classification approach proposed in current paper exploits the method of rule granulation for error reduction and the method of rule consolidation for complexity reduction. The cooperative nature of those methods - the rules are split in a way that makes efficient rule consolidation feasible and rule consolidation itself is capable of further error reduction - is demonstrated in a number of experiments with nine benchmark classification problems. Further complexity reduction, if necessary, is provided by rule compression.

In current study, the ability of computational intelligence methods to tackle modern control problems is under observation. In particular, two computational intelligence techniques belonging to different algorithm families are reviewed, refined and applied to a benchmark fed-batch fermentation process – a linguistic inversion method that designs the controller through automated analysis of information encapsulated in fuzzy linguistic rules of the process model and an evolutionary computation approach that complements its search capabilities with a human critic that guides the optimisation process. The results in terms of productivity measures that compare very favourably to reported results available from literature strongly suggest that in conditions (i.e. presence of noise, parameter variation and randomness) that resemble real-life situations, approximate techniques have an edge over more or less conventional optimisation methods as being more robust and more effective in knowledge discovery.

Fuzzy systems have gained enormous popularity in numerous applications capable of handling uncertainty and complexity of the real world in contrast to conventional techniques. It is the dual nature of fuzzy systems that makes them attractive for modelling and control purposes - in addition to good approximation chat makes them attractive for modelling and control purposes - in addition to good approximation capabilities, the information saved into the fuzzy system can be processed in linguistic terms, much similar to human reasoning. Interpretability that derives from the use of linguistic rulebase is, however, interconnected with the transparency of the fuzzy system, and the latter not being the default property of fuzzy systems, is often ignored by the developers. This paper defines the transparency conditions for standard fuzzy systems and describes the techniques available for obtaining transparent fuzzy systems from experimental data along with examples and comments.