As the complexity of systems increases, their successful modeling becomes a difficult and complex task. Key challenges in system modeling include constructing accurate as well as transparent and highly interpretable models easily comprehended by humans. With this respect, the development of user-centric models endowed with highly interactive interfaces is a highly relevant and timely task.;The objective of our research is to investigate and develop a generalized logic model that is able to achieve a meaningful balance between accuracy and transparency when interacting with users. Such a model can deal efficiently with highly dimensional modeling problems. Fuzzy logic and fuzzy sets are able to cope with linguistic information (information granules) and are therefore compatible with human perception. We exploit the technology of fuzzy neural networks. Such networks combine the superb learning abilities of neurocomputing with the high interpretability aspects associated with fuzzy logic. The design scheme consists of three fundamental phases, namely the design of efficient information granulation mechanisms realized by the interface layout, the formation of learning schemes in the processing core, and the interpretation of model, delivering readable rules back to the user. Several design techniques are presented in the thesis including fuzzy equalization, conditional Fuzzy C-means clustering, particle swarm optimization, gradient-based learning, and network pruning. Experimental studies are reported and the obtained results demonstrate the feasibility and efficiency of the proposed models.
展开▼
