In this paper, we propose an implementation method of kernel adaptive filters by fixing the filter order with lower computational complexity. Kernel adaptive filters are used for adaptive learning of non-linear systems. Although they enable us to estimate non-linear systems, computational load required for implementing the kernel method becomes relatively high. Moreover, the conventional methods require the order of the adaptive filter to be incremented as time n increases. The increment of the filter order results in variation of processing time for updating the filter at each time. These features could cause a problem when we implement them in a system with limited computational resources, such as embedded systems like mobile terminals. We propose, in this paper, a fixed order implementation method of kernel adaptive filters. The proposed method also includes a method to reduce the computational complexity to calculate the Gaussian kernel function. Through the simulation, we show that the proposed method could provide almost same convergence characteristics with less than half of the processing time under certain conditions.
展开▼
