Voting Techniques for Combining Multiple Classifiers

摘要

Algorithms for decision fusion are surveyed and qualitatively compared for theproblem of classification of targets. The methods assume that a number of imperfect classifiers are available, and that these classifiers have enough statistical independence that significant improvement can be made by good combination algorithms. Optimal solutions for this problem require an exact statistical model of the classifiers and the decision space, which are rarely available for real-world problems. Consequently, algorithms must be chosen by intuition and then tested empirically for comparison. Through qualitative comparison, one can reduce the number of algorithms that need to be implemented by eliminating those algorithms that are likely to be weak combiners or to show poor generalization capability. This report surveys candidate algorithms that are likely to show good generalization performance for later empirical evaluation.