The discipline of pattern recognition has seen enormous progress since its beginnings more than four decades ago. Over the years various approaches have emerged, based on statistical decision theory, structural matching and parsing, neural networks, fuzzy logic, artificial intelligence, evolutionary computing, and others. Obviously, these approaches are characterized by a high degree of diversity. In order to combine their strengths and avoid their weaknesses, hybrid pattern recognition schemes have been proposed, combining several techniques into a single pattern recognition system. Hybrid methods have been known about for a long time, but they have gained new interest only recently. An example is the area of classifier combination, which has attracted enormous attention over the past few years.
The contributions included in this volume cover recent advances in hybrid pattern recognition. In the first chapter by H. Ishibuchi and M. Nii, a novel type of neural network architecture is introduced, which can process fuzzy input data. This type of neural net is quite powerful because it can simultaneously deal with different data formats, such as real or fuzzy numbers and intervals, as well as linguistic variables.
The following two chapters deal with hybrid systems that aim at the application of neural networks in the domain of structural pattern recognition. In the second chapter by G. Adorni et al., an extension of the classical backpropagation algorithm that can be applied in the graph domain is proposed. This extension allows us to apply multilayer perceptron neural networks not only to feature vectors, but also to patterns represented by means of graphs. A generalization of self-organizing maps from n-dimensional real space to the domain of graphs is proposed in Chap. 3, by S. Giinter and H. Bunke. In particular, the problem of finding the optimal number of clusters in a graph clustering task is addressed.