Sensitivity analysis of value functional of fractional optimal control problem with application to feedback construction of near optimal controls (Q6166354)

The author considers the differential equation: \((^{C}D^{\alpha }x)(\tau )=f(\tau ,x(\tau ),u(\tau ))\), where \(\tau \in \lbrack 0,T]\) is the time parameter, \(x(\tau )\in \mathbb{R}^{n}\) the state vector, \((^{C}D^{\alpha }x)(\tau )\) the Caputo fractional derivative of order \(\alpha \) defined as: \((^{C}D^{\alpha }x)(\tau )=\frac{1}{\Gamma (1-\alpha )}\frac{d}{d\tau }\int_{0}^{\tau }\frac{x(\xi )-x(0)}{(\tau -\xi )^{\alpha }}d\xi \), \(u(\tau )\in P\) the control vector, and \(P\subset \mathbb{R}^{n_{u}}\) a compact set, with \(n_{u}\in \mathbb{N}\). The function \(f:[0,T]\times \mathbb{R}^{n}\times P\rightarrow \mathbb{R}^{n}\) is supposed to be continuous, Lipschitz continuous with respect to the second variable and to satisfy some boundedness property. A position of the preceding system is a pair \( (t,w(\cdot ))\) consisting of a time \(t\in \lbrack 0,T]\) and a function \( w(\cdot )\in AC^{\alpha }([0,t],\mathbb{R}^{n})\), which is treated as a history of a motion of the preceding system on the time interval \([0,t]\). The set of all such positions \((t,w(\cdot ))\) is denoted by \(G\) and \( G_{0}=\{(t,w(\cdot ))\in G:t<T\}\). For a given initial position \((t,w(\cdot ))\in G_{0}\), the authors considers the set \(X(t,w(\cdot ))=\{x(\cdot )\in AC^{\alpha }([0,T],\mathbb{R}^{n}):x_{t}(\cdot )=w(\cdot )\}\) and the function \(a(\tau \mid t,w(\cdot ))=w(\tau )\), if \(\tau \in [0,t] \), \(a(\tau \mid t,w(\cdot ))=w(0)+\frac{1}{\Gamma (1-\alpha )}\frac{d}{d\tau }\int_{0}^{\tau }\frac{(^{C}D^{\alpha }w)(\xi )}{(\tau -\xi )^{1-\alpha }} d\xi \), if \(\tau \in (t,T]\). An admissible (open-loop) control on the time interval \([t,T]\) is a measurable function \(u:[t,T]\rightarrow P\) and \( \mathcal{U}(t,T)\) is the set of all such controls. The author considers the terminal cost functional \(J(t,w(\cdot ),u(\cdot ))=\sigma (x(T\mid t,w(\cdot ),u(\cdot ))\), \(\forall (t,w(\cdot ))\in G_{0}\), \(\forall u(\cdot )\in \mathcal{U}(t,T)\), where the function \(\sigma \) is Lipschitz continuous. For a given initial position \((t,w(\cdot ))\in G_{0}\), the problem is to find a control \(u(\cdot )\in \mathcal{U}(t,T)\) that minimizes the value of this functional. The author introduces the value functional \(\rho :G\rightarrow \mathbb{R}\) of the preceding optimal control problem and the notion of \(\varepsilon \)-optimal control \(u(\cdot )\in \mathcal{U}(t,T)\) if \( J(t,w(\cdot ),u(\cdot ))=\sigma x(T\mid t,w(\cdot ),u(\cdot ))\leq \rho (t,w(\cdot ))+\varepsilon \). A positional control strategy in the optimal control problem is any mapping \(U:G_{0}\rightarrow P\). The author recalls that a ci-smooth of order \(\alpha \) functional \(\phi :G\rightarrow \mathbb{R} \) is the value functional of the optimal control problem if and only if \( \phi \) satisfies the Hamilton-Jacobi-Bellman equation \(\partial _{t}^{\alpha }\varphi (t,w(\cdot ))+H(t,w(t),\nabla ^{\alpha }\varphi (t,w(\cdot ))=0\), \( \forall (t,w(\cdot ))\in G_{0}\), where \(H\) is the Hamiltonian associated to the optimal control problem and defined as: \(H(\tau ,x,p)=\min_{u\in P}\left\langle p,f(\tau ,x,u)\right\rangle \), \(\forall \tau \in [0,T]\), \( \forall x,p\in \mathbb{R}^{n}\), and the boundary condition \(\varphi (T,w(\cdot ))=\sigma (w(T))\), \(\forall w(\cdot )\in AC^{\alpha }([0,T], \mathbb{R}^{n})\). The author proves that if the value functional \(\rho \) of the optimal control problem is ci-smooth of order \(\alpha \), then a positional control strategy \(U^{o}\) satisfying the condition \( U^{o}(t,w(\cdot ))\in \mathrm{argmin}_{u\in P}\left\langle \nabla ^{\alpha }\rho (t,w(\cdot )),f(t,w(t),u)\right\rangle \), \(\forall (t,w(\cdot ))\in G_{0}\), is optimal in this problem. The first main result of the paper proves under further hypotheses that the value functional \(\rho \) of the optimal control problem is directionally differentiable of order \(\alpha \). Then the author proves that a positional control strategy \(U^{o}\) satisfying the condition \( U^{o}(t,w(\cdot ))\in \mathrm{argmin}_{u\in P}\partial ^{\alpha }\{\rho (t,w(\cdot ))\mid f(t,w(t),u)\}\), \(\forall (t,w(\cdot ))\in G_{0}\), is optimal in the optimal control problem, where \(\partial ^{\alpha }\{\rho (t,w(\cdot ))\mid f(t,w(t),u)\}\) is the derivative of order \(\alpha \) of the value functional \( \rho \) of this problem at the point \((t,w(\cdot ))\) in the direction \( f(t,w(t),u)\). The paper ends with the presentation of three examples.