Parameter optimization of SVR based on DRVB-ASCKF

Abstract

The parameters plays an important role to the performance of support vector regression(SVR). In order to solve the problem of the Parameter optimization for SVR, first, we transform the problem of Parameter optimization into a problem of nonlinear system state estimation, then, we propose a novel algorithm based on Dual Recursive Variational Bayesian Adaptive Square-Cubature Kalman Filter (DRVB-ASCKF), and introduce DRVB-ASCKF to solve it. Considering that the prior statistics noise of a Kalman filter does not agree with its real behavior led to the decrease of the kalman filtering precision, this algorithm assumes that measurement noise variance and process noise variance are unknown in advance, but the function relations between the two kinds of variance are known. This algorithm consists of two iterative processes, during the inner loop using the process noise covariance estimate evaluate measurement noise covariance, and the outer loop using the measurement noise covariance feedback estimate evaluate process noise covariance. Using the DRVB-ASCKF algorithm, we still can get a higher accuracy parameter of SVR when process noise and measurement noise are unknown.