Truncation Error Compensation in Kernel Machines

The analysis and prediction of time series data has played an important role for intelligent systems used in the area of cybernetics and human-machine interaction. Time series prediction is especially important in the case of unreliable communication of data acquired by intelligent systems. Computationally efficient kernel based regression algorithms have allowed for the prediction of non-linear relationships within time series data. In this paper, we present the smooth delta corrected kernel least mean square (SDC-KLMS) algorithm. The SDC-KLMS scales in linear time with the number of samples stored, hence making it computationally efficient. We present a theoretical motivation for our algorithm and we experimentally show how our approach overcomes a limitation imposed by the use of a finite storage buffer. Experiments with simulated, benchmark, and real world data were conducted to verify the accuracy of our algorithm.