A Self-Tuning Filter for Fixed-Lag Smoothing

The problem of estimating a discrete-time stochastic signal which is corrupted by additive white measurement noise is discussed. How the stationary solution to the fixed-lag smoothing problem can be obtained is shown. The first step is to construct an innovation model for the process. It is then shown how the fixed-lag smoother can be determined from the polynomials in the transfer function of the innovation model. In many applications, the signal model and the characteristics of the noise process are unknown. It is shown that it is possible to derive an algorithm which on-line finds the optimal fixed-lag smoother, a self-tuning smoother. The self-tuning smoother consists of two parts: an on-line estimation of the parameters in the one-step ahead predictor of the measured signal, and a computation of the smoother coefficients by simple manipulation of the predictor parameters. The smoother has good transient, as well as good asymptotic, properties.