A Study On Support Vector Machine Implementation In Embedded Control Systems

As a new research achievement of statistical learning theory, Support Vector Machine(SVM) has become the hot research pot over the last more than ten years. Because of its aim to purse the best generalization performance based on small sample set, it can solve most of the problems emerged in the traditional machine learning methods to a large extent, such as nonlinear, local minimum, over-fitting, dimension disaster and other issues. The traits of replacing the empirical risk minimization with the structural risk minimization, convex quadratic programming, mercer kernel function, as well as sparse solutions, make the SVM algorithms have the advantage of simple structure, global optimum and strong generalization ability, etc, and have highlighted their performance advantages on many complex issues from the time when it was proposed. However, the disadvantage of consuming a large storage resources in the SVM training process, especially when the training samples set is big, makes the training speed of the training stage be the bottleneck of practical applications, which restricts the promotion and application of the algorithm to some extent. In real life, a variety of embedded control devices are widely used, which also requires more and better intelligent algorithms can be applied on these platforms efficiently. So, the implementation and application of the SVM algorithms in embedded systems have become the study area of researchers, and they are also the meaningful research directions.Many existing SVM implementations on embedded system platforms focus on the optimization of hardware platforms, including the use of parallel processing units, the use distributed processing and memory cells, and usually exchange the hardware consuming for the training speed upgrade. This article mainly focuses on the study of the embedded control platform implementation of SVM from the software level, and tries to enhance the training speed with the less performance loss, combined the characters of limited resources in embedded control system platforms with the algorithm can be improved. In the paper, we first transmit the training samples from double values to fixed-point or integer values after necessary normalization and fixed-point or integer processing, according to the word length constraint of the target platforms, which could bring a fast training computation speed with certain loss of data precision, then convert the model parameters of standard SVM algorithm to integer values of a range, according to the penalty coefficient and Lagrangian parameters, so that the training stage be to find out the best integer values from0to2k-1as the model parameters; considering the problem of abandon of equal constraint in improved SVM model, the SMO algorithm has no longer been adopted, so the paper adopts the improved SMO algorithm, and presents a new constraint to select the best parameter to be optimized, in this the training speed can be improved.Finally, based on normal PC platform and embedded system toolbox, the paper verified the feasibility of these improvements with, using both artificial data sets and handwritten digital recognition to solve the multi-category classification problems, and analyzed the performances on these platforms.