Robust Optimization

一个科研设想：对于离散的动态规划问题，可以用鲁棒优化类似处理；对于连续的动态规划问题，是否有响应的鲁棒优化理论？

Recall that Linear Optimization (LO) problem is of the form (1) min ⁡ x { c T x + d : A x ≤ b } \min_x\{c^Tx+d:Ax\leq b\}\tag{1} xmin​{cTx+d:Ax≤b}(1) where $x\in {\mathbf{R}}^n$ is the vector of decision variables, $c\in {\mathbf{R}}^n$ and $d\in \mathbf{R}$ form the objective, $A$ is an $m\times n$ constraint matrix, and $b\in {\mathbf{R}}^m$ is the right hand side vector. The stucture of the problem can be given by the following data matrix
$$D=\left[\begin{array}{cc}{C}^T& d\\ A& b\end{array}\right].$$

Definition 1. An uncertain Linear Optimization problem $({\text{LO}}_{\mathcal{U}})$ is a collection (2) { min ⁡ x { c T + d : A x ≤ b } } ( c , d , A , b ) ∈ U \{\min_x\{c^T+d:Ax\leq b\}\}_{(c,d,A,b)\in\mathcal{U}}\tag{2} {xmin​{cT+d:Ax≤b}}(c,d,A,b)∈U​(2) of LO problems (instances) min ⁡ x { c T x + d : A x ≤ b } \min_x\{c^Tx+d:Ax\leq b\} minx​{cTx+d:Ax≤b} of common structure (i.e., with common numbers $m$ of constraints and $n$ of variables) with the data varying in a given uncertainty set $\mathcal{U}\subset {\mathbf{R}}^{(m+1)\times (n+1)}$.

The uncertainty set can be parameterized, in an affine fashion, by pertubation vector $\zeta$ varying in a given pertubation set $\mathcal{Z}$:
$$\mathcal{U}=\left\{\left[\begin{array}{cc}{c}^T& d\\ A& b\end{array}\right]=\underset{\text{nominal data }{D}_0}{\underbrace{\left[\begin{array}{cc}{c}_0^T& d_0\\ A_0& b_0\end{array}\right]}}+\sum _{\ell =1}^L{\zeta }_{\ell }\underset{\text{basic shifts }{D}_{\ell }}{\underbrace{\left[\begin{array}{cc}{c}_{\ell }^T& d_{\ell }\\ A_{\ell }& b_{\ell }\end{array}\right]}}:\zeta \in \mathcal{Z}\subset {\mathbf{R}}^L\right\}\text{\hspace{1em}(3)}$$

Definition 2. A vector $x\in {\mathbf{R}}^n$ is a robust feasible solution to $({\text{LO}}_{\mathcal{U}})$, if it satisfies all realizations of the constraints from the uncertainty set, that is $$Ax\le b,\forall (c,d,A,b)\in \mathcal{U}\text{\hspace{1em}(4)}$$

Definition 3. Given a candidate solution $x$, the robust value $\hat{c}(x)$ of the objective in $({\text{LO}}_{\mathcal{U}})$ at $x$ is the largest value of the “true” objective $c^T+d$ over all realizations of the data from the uncertainty set: $$\hat{c}(x)=\underset{(c,d,A,b)\in \mathcal{U}}{\sup }\left[c^{T}x+d\right]\text{\hspace{1em}(5)}$$

Definition 4. The Robust Counterpart of the uncertain LO problem $({\text{LO}}_{\mathcal{U}})$ is the optimization problem $$\underset{x}{\min }\left\{\hat{c}(x)=\underset{(c,d,A,b)\in \mathcal{U}}{\sup }\left[c^T+d\right]:Ax\le b,\forall (c,d,A,b)\in \mathcal{U}\right\}\text{\hspace{1em}(6)}$$ of minimizing the robust value of the objective over all robust feasible solutions to the uncertian problem.

OBSERVATIONS.
A. The Robust Counterpart of ${\text{LO}}_{\mathcal{U}}$ can be rewritten equivalently as the problem $$\underset{x}{\min }\left\{t:c^T-t\le -d,Ax\le b,\forall (c,d,A,b)\in \mathcal{U}\right\}\text{\hspace{1em}(7)}$$
B. The Robust Counterpart of uncertain LO problem with certain objective min ⁡ x { c T + d : A x ≤ b , ∀ ( A , b ) ∈ U } \min_x\{c^T+d:Ax\leq b, \forall (A,b)\in\mathcal{U}\} xmin​{cT+d:Ax≤b,∀(A,b)∈U} is purely “constraint-wise” construction: to get RC, we act as follows:
(1) preserve the original certain objective as it is, and
(2) replace every one of the original constraints $$(Ax{)}_i\le b_i\iff a_i^{T}x\le b_i\text{\hspace{1em}(8)}$$ ($a_i^T$ is $i$-th row in $A$) with its Robust Counterpart $$a_i^{T}x\le b_i,\forall [a_i,b_i]\in {\mathcal{U}}_i\text{\hspace{1em}(9)}$$ where ${\mathcal{U}}_i$ is the projection of $\mathcal{U}$ on the space of data of $i$-th constraint: U i = { [ a i , b i ] : [ A , b ] ∈ U } . \mathcal{U}_i=\{[a_i,b_i]:[A,b]\in\mathcal{U}\}. Ui​={[ai​,bi​]:[A,b]∈U}. The RC of uncertain LO problem with certain objective remains intact when the original uncertainty set $\mathcal{U}$ is extended to the direct product $\hat{\mathcal{U}}={\mathcal{U}}_1\times \cdots \times {\mathcal{U}}_m$ of its projections onto the spaces of data of respective constraints.

C. If $x$ is a robust feasible solution of (8), then $x$ remains robust feasible when we extend the uncertainty set ${\mathcal{U}}_i$ to its convex hull $\text{Conv}({\mathcal{U}}_i)$.

D. When modeling an uncertain LO problem one should avoid whenever possible converting inequality constraints into equality ones, unless all the data in the constraints in question are certain.

Definition 5. A set $X^{+}\subset {\mathbf{R}}_x^n\times {\mathbf{R}}_u^k$ is said to represent a set $X\subset {\mathbf{R}}_x^n$, if the projection of $X^{+}$ onto the space of $x$-variables is exactly $X$.

Why the representation argument is important? A representation can posses desired properties which are absent in the original problem. Such strategy allows us to focus on a single uncertainty-affected linear inequality: a family { a T x ≤ b } [ a , b ] ∈ U \{a^Tx\leq b\}_{[a,b]\in\mathcal{U}} {aTx≤b}[a,b]∈U​ of linear inequalities with the data varying in the uncertainty set $$\mathcal{U}=\left\{[a;b]=[a^0;b^0]+\sum _{\ell =1}^L{\zeta }_{\ell }[a^{\ell },b^{\ell }]:\zeta \in \mathcal{Z}\right\}\text{\hspace{1em}(10)}$$ naming tractable representation.