Base Vector Learning Mechanism for Fuzzy Model

Fuzzy model based on support vector machine(SVM-based fuzzy model) was proposed in recent years. Although SVM has an excellent generalization performance, it is considered to have lower computation speed, and a large number of support vectors may be found, which leads to a complex fuzzy model with too many rules. To deal with the problem, the paper presents a new approach called base vector learning(BVL) to build fuzzy model. There are two steps in the process of the BVL-based fuzzy modeling. First, the quadratic Renyi entropy is applied to select base vectors, which are used to span a subspace in feature space F. Then, all data are projected onto this subspace where classical algorithms such as classification or regression can be applied. If Gaussian kernel is considered, the structure of BVL is equivalent to fuzzy model. The performance of the proposed learning scheme is illustrated by experiments of classification and regression.