The efficient solution of the (quietly constrained) noisy, linear regulator problem (Q2370738): Difference between revisions

Consider analytic and numerical methods for the pathwise minimization of the functional \[ J(x,u) = \int^b_a f(t,x(t),u(t)) dt + \frac{1}{2} k (x(b)) \] with piecewise-continuous control variable \(u \in H_{a,b}\) (the Cameron-Martin space of absolutely continuous functions with finite quadratic variation), where process \(x\) satisfies the state-independent ordinary stochastic differential equation (SDE of time-shifted Wiener process) \[ dx(t) = u(t) dt + \sigma (t) dW(t) \] driven by a \(m\)-dimensional Wiener process \(W\). The authors discuss some theory of pathwise optimal control problem \[ \min_{u \in H_{a,b}} J(x,u) \] and related numerical algorithms. They claim that their major algorithm (borrowed from deterministic theory) has pathwise global pointwise error of order \(h^{3/2}\) along equidistant partitions with nonrandom mesh size \(h\) in case of such a martingale-type of noise (\(h^2\) in deterministic situation). However, the leading error coefficients are not determined precisely. Essentially, their examples and analysis restrict to the case of the linear regulator, i.e. the optimal control problem with quadratic functionals \(J\) in \(x\) and \(u\), quadratic \(k\) and time-shifted Wiener process (note that the noise \(x\) is governed by a SDE with additive noise, state-independent coefficients and explicitly known strong solution). To prove major results, the theory of deterministic Euler-Lagrange equations and Lagrange multipliers is applied to convert the constrained problem into a nonconstrained one. It may be stressed again that their results are on path-wise optimization with anticipating solutions in contrast to classic moment-wise optimization with nonanticipative controls.