Conditions for generic normality in optimal control

Methods involving semicontinuity and convergence; relaxation (49J45) Nonsmooth analysis (49J52) Optimality conditions for problems involving ordinary differential equations (49K15) Sensitivity, stability, well-posedness (49K40) Nonlinear programming (90C30)

Review text

This paper provides sufficiency conditions for normality of control problems obtained as perturbations of Mayer type problems, extending the results of \textit{R. B. Vinter} [J. Math. Anal. Appl. 96, 153-179 (1983; Zbl 0528.49018); and SIAM J. Control Optimization 21, 235-245 (1983; Zbl 0507.49017)] for ``abnormal control problems. Abnormality means that the maximum principle does not provide additional information, since the multiplier associated with the cost or performance index in the maximum principle turns out to be zero. The control problem (P) considered in this paper is to minimize \(g(x(1))\) subject to \(x\in X_ e\), \(x(1)\in C\) where \(X_ e=\{x\in X|\) \(x(0)\in C_ 0\), \(\dot x(t)=f(t,x(t),u(t))\) a.e., \(u: [0,1]\to {\mathbb{R}}^ m\); \(u(t)\in U(t)\) measurable\} and X is the set of all absolutely continuous functions \(x: [0,1]\to {\mathbb{R}}^ n\). Here C and \(C_ 0\) are closed subsets of \({\mathbb{R}}^ n\), f and g are locally Lipschitz in x and Gr(U) is (Borel \(\times\) Lebesgues) measurable as is \(f(.,x,.)\) for all \(x\in {\mathbb{R}}^ n.\) This problem (P) is considered normal if for every optimal solution \(x^*\), it is possible to choose the corresponding multipliers (p(.),\(\lambda)\) so that \(\lambda =1\). The author defines the concept of ``calmness that uses the lower semicontinuity property of the cost function and implies normality. Specifically, if \[ \phi(a)= \inf\{g(x(1))| \quad x\in X_ e,\quad x(1)\in C+\{a\}\} \] and \(P_ a\) denotes the corresponding minimization problem, then \(P_ a\) is ``calm if \(\phi\) (a)\(\in {\mathbb{R}}\) and \(\liminf_{b\to a} (\phi(b)- \phi(a))/| b-a| >-\infty\). The author proves that if the dynamics of the control problem is linear in the state variable x, then normality holds for all problems obtained by translating the end point set in any direction that is not in a set of measure zero contained in the relative boundary of a convex set. A similar result is obtained in the nonlinear case where translation is replaced by other perturbations. Specifically, if \(Q_ a\) denotes the problem of minimizing g(x(1)) subject to \(x\in X_ e\) and dist\((x(1),C)\leq | a|\), then the problem \(Q_ a\) is normal for almost all \(a\in {\mathbb{R}}^ n\).

Recommendations

Cites work

Cited in

(7)

This page was built for publication: Conditions for generic normality in optimal control

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q1107096)