Choquet integral with regards to a non-additive set function μ is a useful combination tool when we consider the interactions between classifiers. This combination method works very well at the expense of run time and the memory space. This paper introduces samples reduction technology to degrade the complexity of determining the non-additive set functions μ which is determined by genetic algorithm. Reducing samples in training set refers to reduction the outputs of every classifier not the samples themselves. The simulated experiments illustrate that the samples reduction technology can low run time of determining the non-additive set function μ , at the same time, the generalization ability of multiple classifiers combination system is improved mostly.