Optimal linear combination using decision reliability of individual classifiers

This paper presents a new combination method to adaptively assign combination weights according to the decision reliability of each individual classifier, which is estimated from the posterior probabilities of the current sample. The model and training algorithm proposed in this paper are based on the linear combination model and mean squared errors criterion, but can also be extended to other combination methods and criterions. Experimental results on an artificial data set and a practical data set show that the proposed approach can achieve the best classification performance among other linear combination methods