Combining Rough Sets and Neural Network Approaches in Pattern Recognition

Abstract

The paper focuses on problems which arise when two different types of AI methods are combined in one design. The first type is rule based, rough set methodology operating is highly discretized attribute space. The discretization is a consequence of the granular nature of knowledge representation in the theory of rough sets. The second type is neural network working in continuous space. Problems of combining these different types of knowledge processing are illustrated in a system used for recognition of diffraction patterns. The feature extraction is performed with the use of holographic ring wedge detector, generating the continuous feature space. No doubt, this is a feature space natural for application of the neural network. However, the criterion of optimization of the feature extractor uses rough set based knowledge representation. This latter, requires the discretization of conditional attributes generating the feature space. The novel enhanced method of optimization of holographic ring wedge detector is proposed, as a result of modification of indiscernibility relation in the theory of rough sets. Key words: Pattern recognition, neural networks, rough sets, hybrid methods, evolutionary optimization, holographic ring-wedge detectors.