Abstract: A PSO algorithm reduction strategy was proposed to a SVM large-scale training samples by updating the velocity and location of the particles, each particle was corresponding to the status of the training samples, the ideal status included the smallest number of sv, the new training sample has reduced some nsv which arc not effect the SVM classification, so as to reduce the size of the training data sets. A practice of the remote session image classification has proved that the strategy not only has reduced samples, but also enhanced the efficiency of the largcscale data sets training.

Key words: PSO, SVM, Training sample, Large-scale data