A Study Of Transfer Learning Based On Support Vector Machine And Cellular Automata

Transfer learning is suitable for the issues where the source domain and the target domain data are in different data distribution. So it is of more practical application value. Transfer learning methods mainly include transfer learning based on instances, transfer learning based on feature representation, transfer learning based on correlation and transfer learning based on parameters.In this paper, the traditional support vector machine is improved, and we proposed support vector machine method based on different penalty factors. This method is different from the traditional support vector machine, in which each sample has the same penalty factor. By setting the penalty factor for different samples, our method makes the training of the model have strong generalization ability. By using particle swarm optimization algorithm, we search suitable penalty factors for the source domain samples. The penalty factors of source domain samples can reflect the migration ability of the source domain to the target domain, so the improved support vector machine is more capable of transferability.Combining Multiple Attractor Cellular Automata(MACA) with classification and regression tree, the MACA based on classification and regression tree has a good space division ability. At the same time, the MACA based on classification and regression tree has some defects. Especially when the bits of Pseudo Exhaustive Field(PEF) is relatively large, it will cause over fitting. In this paper, the MACA based on classification and regression tree is improved by adaptively adjusting the bits of PEF. Moreover, according to the characteristics of discrete data used by MACA, we give a method to measure the distance of the patterns. This method takes into account the statistical law in different position value after discretization of the samples.This paper finally studies the application of transfer learning based on instances selection in MACA, and the analysis of its problems. Since for the target domain labeled samples are not enough to reflect the distribution of the target domain data, we propose a modified instance selection method. The method makes full use of the ability of dividing the patter space by MACA classification and regression tree, and selects the instances in local mode space.For the improved algorithms proposed in this paper, we have done comparative experiments and verified the validity of the improved algorithms.