System modeling - analysis on transparency, performance and flexibility

System modeling is a prelude task for the analysis in many control and optimization problems. Classification, prediction, and clustering are three typical methods used in system modeling in many science and engineering fields such as data mining, robotics, bioinformatics, and image analysis. And often, these methods have been injected with learning in machine learning and computational intelligence. A viable approach in system modeling is learning from historical data. However, historical data often consist of non-monotonic and inconsistent property in transparency. Because of the shortcomings in previous studies on transparency, performance, and flexibility, three new computational intelligence techniques are proposed and verified in this dissertation for the tasks of classification, prediction, and clustering separately.

At first, the transparency of fuzzy classification systems is studied. A total new transparent linguistic interface generation method for fuzzy systems like fuzzy decision trees is proposed and verified. With the new linguistic interface, the fuzzy decision trees demonstrate both interpretability and improved classification performance.

Secondly, the performance enhancement of neural networks predictors is studied with the traffic flow forecasting problem as the study case. Ensemble methods offer us an approach to improve the available neural network predictors' performance without algorithms revision and complex computation.

Finally, the new fuzzy clustering algorithm named multiple kernel fuzzy c-means is proposed. The multiple kernel fuzzy c-means offers us a flexible vehicle to fusion information from multiple heterogeneous or homogeneous sources. Applications in image segmentation problems demonstrate multiple kernel fuzzy c-means's great advantages and potentials.