Recursive Bayesian estimation using piecewise constant approximations (Q1824597)

The authors consider the following discrete-time stochastic system: \(x_{k+1}=f_ k(x_ k,u_ k)+w_ k\), \(k=0,1,..\). where the system state \(x_ k\), the system transition function \(f_ k\) and the standard white noise \(w_ k\) belong to the space \(R^ n\), while the control parameter \(u_ k\) is in \(R^ m\). The observable process is of the form \(z_ k=h_ k(x_ k,v_ k)\) where \(z_ k,h_ k\in R^ q\), the noise \(v_ k\) is independent of \(w_ k\) and \(v_ k\in R^ p\). Thus the set \(Z_ k=\{z_ i,u_ i:\) \(i=0,1,...,k\}\) is the information avaible at time k. The problem is to find recursively the conditional densities \(p(x_{k+1}| Z_ k)\). Note that generally there is no closed form solution. The well-known linear model is one of a few exceptions. Here the authors obtain a numerical solution by using a simple piecewise constant approximation to the density functions. The proposed algorithm is analyzed in detail. In particular, a bound on the maximum error is derived and stability properties are studied. Useful comparison with other algorithms is given. Finally, the authors apply their algorithm to a system identification problem. The numerical examples are a good illustration of the main results.