Error Bounds for Control Constrained Singularly Perturbed Linear-Quadratic Optimal Control Problems

We present a methodology for bounding the error term of an asymptotic solution to a singularly perturbed optimal control (SPOC) problem whose exact solution is known to be computationally intractable. In previous works, reduced or computationally tractable problems that are no longer dependent on the singular perturbation parameter $\epsilon$, where $\epsilon$ represents a small, non-negative number, have provided asymptotic error bounds of the form $O(\epsilon)$. Specifically, the optimal solution $\bar{V}$ of the reduced problem has been shown to be asymptotically equivalent in $\epsilon$ to the optimal solution $V(\epsilon)$ of the singularly perturbed problem in the sense that $|V(\epsilon)-\bar{V}| =O(\epsilon)$ as $\epsilon \rightarrow 0$. In this paper, we improve on this result by incorporating a duality theory into the SPOC problem and derive an upper bound $\chi_u(\epsilon)$ and lower bound $\chi_l(\epsilon)$ of $V(\epsilon)$ that hold for arbitrary $\epsilon$ and, furthermore, satisfy the inequality $|\chi_u(\epsilon)-\chi_l(\epsilon)| \leq C \epsilon$ for small $\epsilon$, with the constant $C$ determined. We carry out numerical experiments to illustrate the computational savings obtained for the upper and lower bound. In particular, we generate a set of 50 random SPOC problems of a specific form and show that for $\epsilon$ smaller than $10^{-2}$, it becomes faster, on average, to solve for the bounds rather than the SPOC problem and for $\epsilon=10^{-5}$, the computational time for the upper and lower bounds is approximately 20 times faster, on average, than that of the SPOC problem.