Adaptive Stochastic Systems: Estimation, Filtering, And Noise Attenuation

This dissertation investigates problems arising in identification and control of stochastic systems. When the parameters determining the underlying systems are unknown and/or time varying, estimation and adaptive filter- ing are invoked to to identify parameters or to track time-varying systems. We begin by considering linear systems whose coefficients evolve as a slowly- varying Markov Chain. We propose three families of constant step-size (or gain size) algorithms for estimating and tracking the coefficient parameter: Least-Mean Squares (LMS), Sign-Regressor (SR), and Sign-Error (SE) algorithms.The analysis is carried out in a multi-scale framework considering the relative size of the gain (rate of adaptation) to the transition rate of the Markovian system parameter. Mean-square error bounds are established, and weak convergence methods are employed to show the convergence of suitably interpolated sequences of estimates to solutions of systems of ordinary and stochastic differential equations with regime switching.Next we consider problems in noise attenuation in systems with unmodeled dynamics and stochastic signal errors. A robust two-phase design procedure is developed which first estimates the signal in a simplified form, and then applies a control to tune out the noise. Worst-case error bounds are derived in terms of the unmodeled dynamics and variances of the disturbance and measurement errors.