Optimization Of SVM With RBF Kernel And Its Application On Protein Secondary Structure Prediction

There is no a uniform model for the choice of SVM's kernel functions and its parameters. This thesis analyses Grid Search Method (GSM) and Bilinear Search Method (BSM).By improving the two methods, we propose a new parameters optimization method of SVM based on RBF kernel——Bilinear Grid Search Method (BGSM). BGSM combines the advantages of GSM and BSM, appearing to be a more effective learning algorithm. On the condition of keeping the learning accuracy no descending, the method greatly reduces the number of training SVM.BGSM has been checked with UCI data sets. The experiments show that BGSM has better learning performance and learning accuracy than related methods. Application of BGSM to the field of protein secondary structure prediction also obtains better learning accuracy than related algorithms, where moving window and multiple alignment methods are used as encoding methods. By aligning the similarity of unknown sequences and known sequences, we can find whether there is homologous with these sequences, which gives us evolutionary information. By improving GSM and BSM, we propose and implement an optimization algorithm of SVM based on RBF kernel. By improving learning methods and learning strategies, BGSM demonstrates to be better learning performance and learning accuracy compared to other methods.