Time-varying parameters estimation based on Kalman particle filter with forgetting factors

Parameter estimation of time-varying non-Gaussian autoregressive processes is a highly nonlinear problem, which will be more difficult if the functional form of the time variation of the parameters is unknown. In this paper, an efficient particle filter is presented. By integrating Kalman particle filter and the concept of forgetting factors in RLS filter theory, the parameter evolution through time is affected by old and current observations both, and the choice of the unknown parameter distribution is then broadened. Computer simulations proof that the estimation result of this method is more accurate than present approaches in non-linear and non-Gaussian environments, and also the method can suppress the degeneracy of the particles effectively.