Abstract:We consider a residuals-based distributionally robust optimization model, where the underlying uncertainty depends on both covariate information and our decisions. We adopt regression models to learn the latent decision dependency and construct a nominal distribution (thereby ambiguity sets) around the learned model using empirical residuals from the regressions. Ambiguity sets can be formed via the Wasserstein distance, a sample robust approach, or with the same support as the nominal empirical distribution (e.g., phi-divergences), where both the nominal distribution and the radii of the ambiguity sets could be decision- and covariate-dependent. We provide conditions under which desired statistical properties, such as asymptotic optimality, rates of convergence, and finite sample guarantees, are satisfied. Via cross-validation, we devise data-driven approaches to find the best radii for different ambiguity sets, which can be decision-(in)dependent and covariate-(in)dependent. Through numerical experiments, we illustrate the effectiveness of our approach and the benefits of integrating decision dependency into a residuals-based DRO framework.

What problem does this paper attempt to address?

The core problem that this paper attempts to solve is how to construct a residual - based distributionally robust optimization (DRO) model when the uncertainty depends on covariate information and the decision itself. Specifically, the authors propose a new method to deal with two complex factors often encountered in practical decision - making problems: (i) Uncertainty is affected by context/covariate information; (ii) The optimization decision itself has a significant impact on uncertainty. ### Main Research Contents 1. **Problem Background**: - In real - world decision - making problems, decision - makers usually have only limited access to uncertainty realizations or partial knowledge of the underlying distribution. - Traditional stochastic programming (SP) and distributionally robust optimization (DRO) are the two main methods for dealing with uncertainty. However, these methods usually ignore the dependence of the decision on uncertainty, which is crucial in many practical applications. 2. **Proposed Method**: - The authors introduce a residual - based decision - dependent DRO framework (ER - D3RO) to address the above challenges. - Learn the dependence relationship of the random parameter \(Y\) on the covariate \(X\) and the decision \(Z\) through a regression model, and use the residuals obtained during the training process to construct an empirical distribution. - Construct ambiguity sets around this empirical distribution, thereby finding the optimal decision to deal with the worst - case uncertainty. 3. **Theoretical Contributions**: - Provide conditions for statistical properties such as asymptotic optimality, convergence rate, and finite - sample guarantees. - Analyze the performance of solutions under different types of ambiguity sets (such as Wasserstein distance, sample robustness, and same - support ambiguity sets). - Design data - driven algorithms for selecting the radii of ambiguity sets, which can be decision - dependent or - independent, or covariate - dependent or - independent. 4. **Numerical Experiments**: - Verify the effectiveness of the proposed method through numerical experiments, especially under the condition of limited data, and show the benefits of incorporating decision - dependence into the residual - based DRO framework. ### Formula Representation - **Decision - Dependent Conditional Stochastic Programming (DD - CSP)**: \[ v^*(x):=\min_{z\in Z}\mathbb{E}_{Y\sim P}[c(z,Y(x,z))\mid X = x,Z = z] \] - **Residual - Based Empirical Distribution**: \[ \hat{P}_n^{ER}(x,z):=\frac{1}{n}\sum_{k = 1}^n\delta_{\text{proj}_Y(\hat{f}_n(x,z)+\hat{\epsilon}_k^n)} \] - **Wasserstein Ambiguity Set**: \[ \hat{\mathcal{P}}_n(x,z):=\left\{Q\in\mathcal{P}(Y):d_{W,p}(Q,\hat{P}_n^{ER}(x,z))\leq\xi_n(x)\right\} \] In this way, the paper not only extends the existing residual - based DRO framework but also provides a complete set of theoretical and practical tools for dealing with decision - dependent uncertainty.

Residuals-Based Contextual Distributionally Robust Optimization with Decision-Dependent Uncertainty

Improving Robust Decisions with Data

Distributionally Robust Optimization with Multimodal Decision-Dependent Ambiguity Sets

Multistage Robust Mixed-Integer Optimization under Endogenous Uncertainty

Distributionally robust optimization with decision dependent ambiguity sets

Achieving Robust Data-driven Contextual Decision Making in a Data Augmentation Way

Tractable Distributionally Robust Optimization with Data

Distributionally Robust Regret Minimization

Distributionally Robust Optimization Under a Decision-Dependent Ambiguity Set with Applications to Machine Scheduling and Humanitarian Logistics

Multi-stage Distributionally Robust Convex Stochastic Optimization with Bayesian-type Ambiguity Sets

Distributionally Robust Optimisation with Bayesian Ambiguity Sets

Distributionally Robust Decision Making Leveraging Conditional Distributions

Online Optimization and Ambiguity-based Learning of Distributionally Uncertain Dynamic Systems

Adjustable Distributionally Robust Optimization with Infinitely Constrained Ambiguity Sets

Mathematical Foundations of Robust and Distributionally Robust Optimization

Distributionally robust optimization for sequential decision-making

Computationally Efficient Approximations for Distributionally Robust Optimization Under Moment and Wasserstein Ambiguity

Distributionally Robust Optimization with Confidence Bands for Probability Density Functions

From Data to Decisions: Distributionally Robust Optimization Is Optimal

Robust Data-Driven Decisions Under Model Uncertainty