Perturbation of multivariable linear quadratic systems with jump parameters (Q4540231)

The authors consider a perturbation control problem for a class of continuous-time linear-quadratic systems whose position \(x(t)\in \mathbb{R}^p\) at time \(t\) is described as follows: \[ \frac{dx_t} {dt}=A (\theta_t) x_t+B(\theta_t) u(t),\;x(0)=x_0. \] Here \(A\) and \(B\) are matrices, \(u\) the control function taking values in \(\mathbb{R}^r\). The random changes from one structure of the system to another one are governed by the finite-state Markov process \(\theta_t\), \(t\geq 0\), whose infinitesimal generator matrix \(\Lambda=(\lambda_{ij})\) depends on a small perturbation parameter \(\varepsilon\). The control \(u\) is of a Markovian type, i.e. \(u(t)=\gamma(t,x_t,\theta_t)\), the running cost function is given, and the goal is to optimize a discounted expected cost function. By using aggregation techniques to reduce the dimension of the Markov chain \(\theta(t)\), the authors are able to study a coupled set of Riccati equations to determine the optimal value and the control. The main result is the derivation of Taylor expansions for these two quantities with respect to the perturbation parameter \(\varepsilon\). An algorithm is presented for computing the terms of the series.