Fusing support vector machines and soft computing for pattern recognition and regression.

摘要

Soft computing is ideally suited for dealing with vague, uncertain and complex information. However, the classical fuzzy systems, neural networks and neuro-fuzzy systems usually suffer from severe drawbacks such as convergence problem, generalization problem, the "curse of dimensionality," local optimum, etc. Support vector machine (SVM), a novel learning method derived within the statistical learning theory, usually achieves superior performance compared with traditional soft computing methods. The purpose of the research is to study the relationships between SVMs and soft computing methods and fuse them in both directions in order to preserve their respective advantages. The research in this dissertation is threefold as follows.; First, the theoretical connections are constructed between SVM learning/kernel methods and two typically soft computing technique---fuzzy additive models and fuzzy adaptive network. Most of the possible fuzzy systems qualified in such connections are analyzed. A unified framework of support vector learning based fuzzy systems is proposed, in which SVM automatically identifies the optimal fuzzy model structure.; Second, a series of new support vector fuzzy systems (SVFSs) based on fuzzy additive models with positive semi-definite fuzzy basis functions are proposed. The proposed SVFSs are also extended to include the non-Mercer kernel based systems. These SVFSs can efficiently overcome the drawbacks of the classical soft computing methods, and strike a fine balance between the model complexity and the approximating accuracy. In addition, by applying SV learning to classical neuro-fuzzy systems, support vector fuzzy adaptive networks (SVFANs) are proposed. The proposed SVFANs combine the superior learning power of SVM and the efficient human-like reasoning and adaptation capacity of fuzzy adaptive network in handling uncertain information.; Third, the concept of fuzzy set theory is incorporated into SVM. A fuzzy kernel is constructed based on fuzzy similarity measure between training data. This strategy is compared with another strategy in which the fuzzy kernel is estimated from the symmetric triangular fuzzy membership functions. The proposed fuzzy kernels and their corresponding fuzzy SVMs are shown to be useful in dealing with the information in a vague, uncertain and complex environment as well as preserving all the advantages of SVMs.