Monte Carlo minimization for one step ahead sequential control (Q4938852)

The one step ahead sequential control for the model: \[ y_t = h(\omega_t,z_t,\beta) + e_t, \qquad t=1,2,\ldots , \] aims to determine the control \(\omega_t\) that minimizes a predictive expected loss (PEL) based on the data up to time \(t\). The error terms \(e_t\) are i.i.d. with variance \(\sigma^2\). The vector \(\beta\) and the variance \(\sigma^2\) are unknown parameters. Typically, there are no closed form solutions to this problem. The authors propose a minimization procedure based on ordinary Monte Carlo or Markov Chain Monte Carlo. They prove convergence theorems and asymptotic normality. The results are similar in spirit to those available for maximum likelihood and \(M\)-estimation but the theory developed here requires weaker regularity conditions due to the use of epiconvergence. Examples are given for the control of dynamic models with regression and autoregression-type terms using various loss functions. For the entire collection see [Zbl 0919.00063].