Distributionally Robust MILP Framework

• DRO-MILP frameworks are robust optimization techniques that incorporate finite sample data and ambiguity sets to handle objective uncertainty in MILP problems.
• They utilize geometric metrics like Wasserstein and divergence measures such as Kullback–Leibler to define plausible distributions and enable tractable single-level MILP reformulations.
• These methods deliver robust decision-making with strong out-of-sample performance and statistical guarantees, applicable to various combinatorial optimization challenges.

Distributionally robust optimization (DRO)-MILP frameworks provide a principled methodology for solving mixed-integer linear programming (MILP) problems under objective function uncertainty using finite data samples. These frameworks construct ambiguity sets representing probabilistic uncertainty, enabling robust prescriptions with statistical guarantees. Contemporary DRO-MILP paradigms utilize either geometric metrics (e.g., Wasserstein distance) or information-theoretic divergence measures (e.g., Kullback-Leibler relative entropy) to define the space of plausible distributions, facilitating tractable reformulations and strong out-of-sample performance.

1. Problem Structure and Fundamental Principles

Let a decision-maker choose $x \in X$, with $X$ a poly-mixed-integer set (i.e., $$X \subseteq \mathbb{R}^{n_1}_+ \times \mathbb{Z}^{n_2}_+$$ or $$X = \{ x: A_{\text{cont}} x_{\text{cont}} + A_{\text{bin}} x_{\text{bin}} \leq b, x_{\text{cont}} \in \mathbb{R}^n, x_{\text{bin}} \in \{0,1\}^m \}$$), to minimize worst-case expected loss under uncertainty in the objective function parameters.

Given a random cost vector $c \in \mathbb{R}^n$ with unknown distribution, only a finite training data set is available, often subjected to incomplete or noisy observations. Two principal uncertainty sources are addressed:

• Data uncertainty: The true value of each observed cost sample $\hat c^{(k)}$ lies within a sample-wise polyhedron $\mathcal{S}_k = \{ c: B^{(k)} c \leq b^{(k)} \}$, subsumed by a global support $\mathcal{S}_0$ (Ketkov, 2023).
• Distributional uncertainty: For a fixed $\widehat C = (\hat c^{(1)}, \dots, \hat c^{(K)})$, the ambiguity set consists of distributions $Q$ at prescribed divergence (Wasserstein or relative-entropy) from the empirical distribution.

A canonical three-level optimization model is then formulated: $$\min_{x \in X} \max_{\hat c^{(k)} \in \mathcal{S}_k\ \forall k} \max_{Q \in \mathcal{Q}(\widehat C)} \mathbb{E}_{c \sim Q}[\ell(x, c)],$$ where $\ell(x, c)$ is typically biaffine: $$\ell(x, c) = c^\top T x + t_1^\top x + t_2^\top c + t_0.$$

2. Ambiguity Set Construction and Uncertainty Models

Ambiguity sets are critical to DRO. Two principal constructions are supported:

• Wasserstein metric ambiguity set: Defines neighborhood around the empirical distribution $$\hat Q_K = \frac{1}{K} \sum_{k=1}^K \delta_{\hat c^{(k)}}$$ via the Wasserstein–1 distance w.r.t. the $\ell_1$–norm:

$$W^1(Q, Q') = \inf_{\pi \in \Pi(Q, Q')} \int_{\mathcal{S}_0 \times \mathcal{S}_0} \| c - c' \|_1 \; \pi(dc, dc').$$

The ambiguity set is then $$\mathcal{Q}(\widehat C) = \{ Q: W^1(Q, \hat Q_K) \leq \epsilon_K \}$$ (Ketkov, 2023).

• Component-wise relative-entropy ambiguity set: For each component $j$, independent samples $c_{j,1}, \dots, c_{j,T_j}$ are observed; empirical marginals $\hat Q_j$ are constructed. The ambiguity set is

$$\mathcal{P} := \{ Q = \otimes_{j=1}^n Q_j : D_{KL}(Q_j \| \hat Q_j) \leq \epsilon_j\ \forall j \},$$

with $D_{KL}$ the Kullback-Leibler divergence. The inner supremum over $Q$ decouples into $n$ univariate convex programs (Ketkov et al., 2021).

Data uncertainty is modeled polyhedrally and admits special cases:

• Interval uncertainty: $\ell_a^{(k)} \leq c_a \leq u_a^{(k)}$ for intervals.
• Semi-bandit and bandit feedback: Partial exact observations or aggregate costs; see computational treatments below.

3. Single-Level MILP Reformulations

Both frameworks admit tractable reformulations contingent on the structure of the loss function and ambiguity set.

• Wasserstein DRO MILP: For biaffine $\ell(x, c)$, the three-level problem is reformulated as a single-level MILP by dualizing the inner maximization. Using Esfahani–Kuhn duality, one obtains:

$$\max_{Q: W^1(Q, \hat Q_K) \leq \epsilon_K} \mathbb{E}_Q[\ell(x, c)] = \min_{\lambda \geq 0, s \in \mathbb{R}^K} \left\{ \lambda \epsilon_K + \frac{1}{K} \sum_{k=1}^K s_k \right\}$$

subject to constraints derived from LP duality on sample-wise data uncertainty. The outer maximization over $\widehat C$ and the minimization over dual variables are exchanged via Sion's min-max theorem, resulting in the final MILP:

$$\min_{x \in X, \lambda \geq 0, \nu^{(k)} \geq 0, \gamma^{(k)} \geq 0} \left\{ \lambda \epsilon_K + \frac{1}{K}\sum_{k=1}^K b^{(0)\top} \nu^{(k)} + \sum_{k=1}^K b^{(k)\top} \gamma^{(k)} + t_1^\top x + t_0 \right\}$$

with coupling and polyhedral constraints indexed by $k$ (Ketkov, 2023).

• Relative-entropy DRO MILP: The inner supremum for each $j$ reduces to a convex program to compute $c^*_j(\epsilon_j)$:

$$c^*_j(\epsilon_j) = \min_{\beta_j \geq \bar z_j} \left\{ \beta_j - e^{-\epsilon_j} \prod_{k=1}^{d_j} (\beta_j - z_{j,k})^{\hat q_{j,k}} \right\}$$

The overall DRO problem is then a deterministic MILP:

$$\min \sum_{j=1}^n c^*_j(\epsilon_j) x_j\quad \text{subject to}\quad x \in X.$$

No new integer variables are introduced; off-line convex minimizations are required for each $c^*_j$ (Ketkov et al., 2021).

Special cases (interval or bandit feedback) allow further reductions to standard MILPs or closed-form enumerations.

4. Statistical Guarantees and Performance Metrics

Both approaches yield robust prescriptions with finite-sample and asymptotic guarantees:

• Prediction & prescription guarantees: For chosen radii $\epsilon_j$, the DRO predictor $\sum_j c^*_j(\epsilon_j) x_j$ is Pareto-undominated among all prediction rules with exponential out-of-sample disappointment guarantees and is strongly optimal under affine-support conditions (Ketkov et al., 2021).
• Asymptotic rates: Choice $$\epsilon_j = (1/T_j)[d_j \ln(T_j + 1) + r T_{\text{min}} - \ln \delta_j]$$ enforces exponential decay of underestimation probability with rate $r$ as $T_{\text{min}} \to \infty$ (Ketkov et al., 2021).
• Out-of-sample performance metrics: The nominal relative loss metric

$$\rho(x) = \mathbb{E}_{Q^*}[\ell(x, c)] / \min_{x'} \mathbb{E}_{Q^*}[\ell(x', c)]$$

is deployed to quantify robustness (Ketkov, 2023).

A plausible implication is that these guarantees enable tight control of conservatism and disappointment risk, particularly for practitioners utilizing finite data streams in sequential decision environments.

5. Computational Aspects and Special Structures

Computational studies examine tractability and efficiency:

• The presented MILP reformulations are of comparable size and complexity to the nominal MILPs; interval and semi-bandit special cases reduce to solving standard MILPs; bandit feedback structures admit closed-form enumeration using the sample average costs (Ketkov, 2023).
• For sorting, shortest-path, and maximum-coverage problems with sample sizes up to $K=100$ and dimensions $n$ up to $70$, the proposed MILPs are solved in seconds for moderate $K$ and exhibit LP relaxation gaps of $1$–$5\%$ for bandit feedback (Ketkov, 2023).
• Sparsity in the decision variable $x$ (e.g., small $h$ in SPP) beneficially reduces MILP difficulty; cases with dense coverage grow computationally costlier but remain tractable for moderate problem sizes.

Computational algorithms proceed in two stages: precompute problem-specific worst-case coefficients (by solving small convex programs), then solve a single instance of the nominal MILP (Ketkov et al., 2021).

6. Significance, Applications, and Practical Implications

DRO-MILP frameworks enable robust decision-making under compound data and distributional uncertainty in stochastic combinatorial optimization. Applications include:

• Sorting and selection problems
• Shortest-path determination in layered graphs
• Maximum coverage optimization in bipartite collections
• Knapsack instances with variable sample availability

Numerical results confirm:

• Rapid convergence of DRO solutions to the true-optimal decision as sample size increases (relative loss $\to 1$).
• Significant reductions in conservatism compared to classical robustification approaches (e.g., Hoeffding bounds).
• Adaptivity to real-world sampling phenomena (such as uneven observation counts across problem components).

This suggests broad applicability for operations research practitioners, particularly in data-driven or bandit-like sequential settings.

7. Connections to Related Research and Methodological Remarks

The presented DRO-MILP approaches generalize and strengthen traditional robust optimization protocols by introducing statistical optimality and tractable reformulations. Key methodological advances include decoupled inner supremum programs for component-wise relative entropy balls and the reduction of three-level min–max–max problems to single-level MILPs via duality, min-max exchange, and convex optimization.

A plausible implication is that future extensions may consider non-biaffine losses, continuous data feedback, or adaptive ambiguity set tuning to further enhance practicality and scope. The framework’s reliance on tractable reformulations and off-the-shelf MILP solvers reinforces its deployability across diverse application domains.

---

Citations:

(Ketkov, 2023): A study of distributionally robust mixed-integer programming with Wasserstein metric: on the value of incomplete data (Ketkov et al., 2021): On a class of data-driven mixed-integer programming problems under uncertainty: a distributionally robust approach