ON COMBINING CLASSIFIERS

We develop a common theoretical framework for combining classifiers wh ich use distinct pattern representations and show that many existing s chemes can be considered as special cases of compound classification w here ail the pattern representations are used jointly to make a decisi on. An experimental comparison of various classifier combination schem es demonstrates that the combination rule developed under the most res trictive assumptions-the sum rule-outperforms other classifier combina tions schemes. A sensitivity analysis of the various schemes to estima tion errors is carried out to show that this finding can be justified theoretically.