Pattern classification. Neuro-fuzzy methods and their comparison

Publication date 17 September 2001

neuro-fuzzy methods pattern classification

Introductory exposition (textbooks, tutorial papers, etc.) pertaining to computer science (68-01) Learning and adaptive systems in artificial intelligence (68T05) Pattern recognition, speech recognition (68T10)

Review text

Multilayer neural networks and fuzzy systems are widely applied to pattern classification and function approximation. Although the two approaches are proved to be mathematically equivalent in that they are convertible, they have their advantages and disadvantages. Multilayer neural networks can learn from data: just gather data and train the network using the back-propogation algorithm; than we can get the classifier or the function approximator that we want. However, if the trained network does not perform as we have expected, we need to retrain the network changing the network size, the number of hidden neurons, or adding new training data. In developing fuzzy systems one need to divide the input space in advance. Thus as the number of input variables increases, the number of rules to be defined increases exponentially. This is called the curse of dimensionality. In addition, since the input space is divided into subregions that are parallel to input axes, performance of fuzzy systems is usually inferior to that of multilayer neural networks. This book consists of two parts: the first part from Chapter 1 to Chapter 12 discusses pattern classification and the second part from Chapter 13 to Chapter 16 discusses function approximation. Chapter 1 explains the pattern classification task and clarifies the problems to be considered in developing pattern classification systems. Chapter 2 discusses the neural network classifier and explain the conventional back-propagation algorithm and the training algorithm based on the synthesis principle of the neural network classifier. Chapter 3 discusses the support vector machines for pattern classification which are based on statistical learning theory. The input space is mapped into a Hilber space called feature space. In the feature space, the classifier is constructed so that the training data of one class are separated by the hyperplane with the maximum margin from those of the remaining classes. Training results in solving a quadratic optimization problem. Conventional fuzzy classifiers divide the ranges of the input variables into several sub-intervals that correspond to fuzzy sets, and define a one-dimensional membership function for each interval. In Chapter 4, the author defines one-dimensional membership functions. Then he extends them to multi-dimensional membership functions. In generating fuzzy rules, overlaps between classes need to be resolved. This can be done before and after rule generation, and during rule generation. Chapter 5 discusses static generation of fuzzy rules with different type of fuzzy rules and their training architecture. Chapter 6 discusses clustering techniques that are used to divide class data into clusters and Chapter 7 discusses how to tune membership functions of the fuzzy rules that are generated by static rule generation. Chapter 8 discusses robust pattern classification using ellipsoidal fuzzy rules when outliers are included. Chapter 9 discusses dynamic generation of two types of hyperbox fuzzy rules. One method resolves overlaps between hyperboxes of different classes by expanding or contracting hyperboxes. The other method resolves overlaps by defining overlapping regions as an inhibition hyperbox. Chapter 10 compares classification performance and the computation time of the fuzzy classifiers discussed in this book. Performance of pattern classification depends heavily on the input features used. Chapter 11 discusses the methods for extracting and selecting important features from the given input features. Feature extraction extracts important features by linear or nonlinear transformation of the original input features. In developing a pattern classification system, usually we prepare two data sets: one for training and the other for testing to check generalization ability. Since the training data needs to be unbiased representation of the events, the characteristics of the two sets need to be similar. To realize this, random division of the data is used. In Chapter 12, the similarity of two data sets is measured by the similiarity of centers and covariance matrices of the two. Then the division algorithm is proposed and the effectiveness of the method for two benchmark data sets is shown. Chapter 13 is a survey of function approximation methods using neural networks, conventional fuzzy systems, and fuzzy systems with learning capability. Chapter 14 discusses function approximators with hyperbox fuzzy rules and two types of fuzzy function approximators with ellipsoidal fuzzy rules: FACG and FALC. Chapter 15 discusses fuzzy rule generation by preclustering and postclustering for FACG and FALC and tuning of membership functions. Chapter 16 discusses robust fuzzy function approximation when outliers are included.

Recommendations

Cited in

(16)

Describes a project that uses

Uses Software

SVMlight

This page was built for publication: Pattern classification. Neuro-fuzzy methods and their comparison

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q2740522)