Recently, in the fields of pattern recognition and machine learning, the technology of information fusion develops very fast and expands its wide application. There are three stages in information fusion: data fusion, character fusion and decision fusion. A lot of approaches to classifier fusion are decision fusion or experts mixture. In many realistic cases, multiple classifiers are not independent but have interaction. In this thesis, we use the Choquet fuzzy integral operator to fuse the different neural network classifiers which have been trained in advance. The calculation of the Choquet integral can be transferred into the linear combination about the fuzzy measure, which is differentiable. This quality allows us to use the standard optimization technology to determine the fuzzy measure. In this thesis, we use both the linear programming and quadratic programming to determine the fuzzy measure, these methods are tested in a data set. The experiments show that the fusion accuracy based on the fuzzy measure which is determined by the quadratic programming is higher than the accuracy based on the fuzzy measure which is determined by the linear programming. There are almost no difference in time and space complexity.
|