基于分类器判决可靠度估计的最优线性集成方法

多分类器集成的目标是充分利用每一个分类器的长处，既要发挥它们的最佳性能，又能克服单个分类器的弱点，以达到最优的系统识别性能，该文提出一种基于发类器判决可靠度估计的多分类器集成方法，利用各分类器对当前输入样本的判决兵验概率，实时估计它们的分类判决可靠度，并指导集成权重系数的分配，该方法能够使集成权重更灵活地适应不同的输入变化，充分利用每个分类器当前的识别能力，从而获得更好的集成效果，该文结合最小均方误差准则（MSE）下的最优线性集成方法给出了权重模型和训练算法，并与其它的线性集成模型作了比较，实验表明，不论在最优化准则下的最终误差还是在实际识别性能上，作者提出的方法都取得了最好的效果。

An Optimal Linear Combination Method by Evaluating the Reliability of Individual Classifiers

This paper presents a new optimal linear combination(OLC) method which adaptively assigning combination weights according to the current decision reliability of each individual classifier. The current reliabilities are evaluated from the posterior probabilities of the current input sample to make sure that the decision reliability represents the classifier's current effectiveness. So the combination system could fully utilize effectiveness of every classifier and acquire best performance if the classifier with higher decision reliability has a higher weight. Such a new method is believed to be more flexible and more accurate than only one set of fixed weights were used when classifying different samples. The experimental results on simulation data and online handwritten Chinese character data sets has shown that the method achieves the best classification performance among other linear combination methods. The analysis and training algorithm of the new method proposed in this paper are based on the OLC method and mean squared errors criterion, but can also be extended to other combination forms and criterions.