Designing fuzzy controllers involves both identifying suitable fuzzy dicretizations of the input and output spaces (number of fuzzy terms and related membership functions) and drawing effective rule bases (hopefully compact and complete). Learning from examples is a very efficicnt way for acquiring knowledge. Moreover, providing reliable examples is easire than directly ddesigning and coding a control strategy. Supervised learning is therefore appealing for automatically building control systems. It may also be used as quick start-up for applications requiring further on-line learning (through reinforce-based techniques or processing of further appropriate example). Supervised techniques for building rule bases form suitable examples generally rely on pre-defined fuzzy discretizations of the input and output spaces. This paper proposes the use of an ART-based neural architecture for identifying, starting from examplex, an appropriate set of fuzzy terms and associated membership functions. These data are then used by an ID3-based machine learning algorithm for building fuzzy control rules. The ART framework provides fast convergence and incremental building of classes, gracefull accounting for the integration of new sample data. The whole chain (the neural architecture for building fuzzy discretizations and the machine learning algorithm for drawing fuzzy rules) has been proved on examples provided by several operators, with different skills, driving a real vehicle along the right-hand wall in an indoor environment. The obtained results are shown, discussed and compared with the performance of controllers using human difined fuzzy partitions.
展开▼
