Knowledge integration in a multiple classifier system

This paper introduces a knowledge integration framework based on Dempster-Shafer's mathematical theory of evidence for integrating classification results derived from multiple classifiers. This framework enables us to understand in which situations the classifiers give uncertain responses, to interpret classification evidence, and allows the classifiers to compensate for their individual deficiencies. Under this framework, we developed algorithms to model classification evidence and combine classification evidence from difference classifiers, we derived inference rules from evidential intervals for reasoning about classification results. The algorithms have been implemented and tested. Implementation issues, performance analysis and experimental results are presented.