Support vector machine data reduction for Direct Filter

In this paper, an approach for speeding up Direct Filter (DF), which is a method to compute optimal filters for nonlinear systems, is proposed. DF is a nonparametric state estimation method, where the filter is designed from the plant data directly. Direct filtering employs a clever strategy of averaging two nonlinear functions trained from the training data to obtain optimal filters for a restricted class of nonlinear systems, with the assumption that plenty of data is available. The filter can then be used to predict future states. While a clever idea, its nonparametric nature makes it slow. The authors propose to use Support Vector Machine (SVM) to suppress the size of data. Since SVM learns support vectors, the authors propose that the support vectors are the actual data of interest, and thus should be used for the nonlinear function estimation employed by the direct filter. Experimental results show good speedups with minimal degradation.