10-Minute Flash Talk Competition

The 10-Minute Flash Talk Competition offers a fast-paced platform for PhD and graduate students to present their research projects to a broader audience. Selected participants will deliver brief, engaging presentations, with standout talks receiving cash prizes!

Accepted Flash Talks

Multistage Adaptive Robust Optimization for Long-term Energy Infrastructure Resilience Planning

Wei Gu, Carnegie Mellon University

In long-term resilience planning, energy infrastructure operators often face a sequence of irreversible and long-lasting strategic investment decisions under sequential revelation of future uncertainties, spatiotemporal resource constraints, and system evolution. Motivated by these challenges, we present a multistage adaptive robust optimization with multiplicative objective (MS-ARO-MO) framework for long-term energy infrastructure resilience planning that dynamically reallocates protection and investment as new information unfolds, subjecting to spatiotemporal resource limitation across locations and over time. The proposed MS-ARO-MO framework includes an objective function with multiplicative structure which captures the coupling effects between uncertainty and a sequence of irreversible long-lasting decisions that are pervasive in energy infrastructure resilience applications. To solve the MS-ARO-MO modeling framework efficiently, we develop a new class of exact and approximate network flow formulations, and establish the connection between MS-ARO-MO and traditional static robust optimization as well as two-stage robust optimization. Computational experiments demonstrate that MS-ARO-MO substantially outperforms various benchmark models by enabling earlier identification of high-risk locations, avoiding myopic over-hardening, and yielding more reliable long-term decisions. Our results highlight the importance of adaptive, risk-aware optimization for long-term energy infrastructure resilience in the face of compounding climate and operational uncertainties.

Robust Chance-Constrained Optimization using a Continuous Parameter Space Wasserstein-2 Ambiguity Set of Gaussian Mixtures

Shibshankar Dey, Northwestern University

We study distributionally robust (DR) chance-constrained optimization under Gaussian mixture uncertainty. Finite-support DR (FDR) stress-tests empirical mixture supports but can miss nominal-parameter misspecification. We address this by developing an ambiguity set that uses the Bures–Wasserstein metric over probability measures and lets the worst-case distribution endogenously determine both how many mixture components receive mass and where their parameters lie within a continuous support. This continuous support DR (CDR) model significantly generalizes a previously known finite-set-based approach from ML community. A case study using energy-allocation problem in an electric vehicle charging station demonstrates the framework’s practical value in achieving any reliability targets via a strong-duality-based semi-infinite reformulation and an adaptive cutting-surface algorithm. A block-alternating local-search procedure in the proposed cutting surface approach is used to identify new Gaussian distributions added to the current pool. Across service-level targets $\theta\in\{0.95,0.97,0.99\}$ and Wasserstein radii $\rho\in\{0.001,0.005,0.01\}$, the results show superior out-of-sample chance constraint satisfaction when compared with no robustification or a finite-support-based robustification of a nominal GMM-based chance constraint. In the out-of-sample testing, FDR fails to attain the prescribed target probability for every tested $(\theta,\rho)$ pair. In contrast, CDR exhibits a consistent $\rho$-based improvement in out-of-sample chance constraint satisfaction (OSS) probability as \(\rho\) increases under \(\pm5\%\) or \(\pm10\%\) support uncertainty allowance from the mean of the nominal Gaussians. A detailed analysis of the generated solutions shows that the CDR model makes structural changes to the energy allocations, whereas the FDR model allocations are close to those from the nominal model. Thus, CDR shifts robustification from finite-support nominal-mixture stress testing to protection against structural demand misspecification.

Speed of Intervention in Algorithmic Markets: Controlling Collusion and Stability

Tong Xie, University of Chicago Booth School of Business

Algorithmic pricing is increasingly prevalent in online marketplaces, raising concerns that algorithms may learn to collude even without explicit communication. This paper investigates how the design of learning algorithms shapes the emergence of collusion and the effectiveness of platform interventions intended to mitigate it. We model the interaction between competing sellers as a repeated Prisoner’s Dilemma where firms deploy a value-based class of learning algorithms that we refer to as Q-learning Reinforcers. We study an intervention where the platform steers demand to reallocate exposure toward sellers choosing more competitive actions. Our analysis shows that the effectiveness of this intervention depends critically on both its level and its speed of implementation. While setting the intervention level too high can destabilize the system, the timing of the intervention also matters: sudden interventions may induce instability through a subcritical Hopf bifurcation, thereby resulting in persistent market oscillations. In contrast, gradual implementation allows the learning dynamics to track the moving stable equilibrium, guiding the market toward a more competitive stable outcome without triggering oscillations. Our results highlight that successful market regulation requires choosing not only how much to intervene, but also the speed of intervention.

Utilizing External Predictions under Selective Feedback

Hongyu Chen, MIT

Estimating population quantities such as mean outcomes from user feedback is fundamental to platform evaluation and social science, yet feedback is often missing not at random (MNAR): users with stronger opinions are more likely to respond, so standard estimators using only observed outcomes are biased and the estimand is not identified without additional assumptions. In this paper, we develop a partial identification framework for estimating the population mean in the existence of a special class of auxiliary variable -- Weak Shadow Variables. We define weak shadow variables to be imperfect proxies of the missing outcomes that is independent of the missingness once conditioned on the outcome and covariates. Specifically, they need not to satisfy a completeness condition required by classical shadow-variable methods, which naturally adapts to outcome predictions from pretrained models, including large language models (LLMs). Under this setup, the identification region is an interval with endpoints solve by a linear program. In finite samples, estimation and inference can be problematic because a direct plug-in estimator can be infeasible. Thus, we propose a local penalized estimator that is feasible in finite sample and achieves $\sqrt{n}$ convergence rate. We also propose a subsample bootstrap inference to construct confidence intervals for the identification region. In simulations and semi-synthetic experiments on real customer-service dialogues, we find that our proposed method has superior performance compared to other classic MNAR methods even with simple binary weak shadow variables.

Biobjective Pareto Paths: Analysis and Algorithm

Guanting Wu, Carnegie Mellon University

We aim at obtaining the biobjective properly Pareto optimal set via studying the optimal solutions to a series of scalarized biobjective optimization problems. We focus on the linearly coupled nonsmooth problems, which subsumes the lasso regression, the grouped lasso, and the regularized support vector machine (SVM). For any positive weighting scalar of the two objectives, we characterize the continuous function, dubbed as the Pareto path, that maps the scaler to the optimal solution. Theoretically, we bypass the nonsmoothness of the primal objectives via eliciting a hidden smoothness of the augmented Lagrangians. Then, we invoke an implicit function theorem to select a Lagrange multiplier path, dubbed as a shadow path, that (exactly) corresponds to the Pareto path over some interval of the weighting parameter. Computationally, we design an Euler-Newton continuation algorithm that follows a shadow path to (approximately) recover the properly Pareto optimal set. Across different applications, our algorithm shows superior numerical performances over benchmarks, especially in high-dimensional regime, and our analysis provides new insights into the exact path of the grouped lasso.

Homogeneous Quadratically Constrained Quadratic Programming: Stronger Approximation Bounds and Exact Methods

Haoyun Deng, Georgia Institute of Technology

Homogeneous quadratically constrained quadratic programming (HQCQP) is a challenging class of nonconvex optimization problems. Our approach exploits the Lagrangian relaxation and its dual information to develop both a new approximation algorithm and an efficient exact solution framework. We solve the Lagrangian relaxation using a bundle method and combine it with Gaussian rounding to construct high-quality feasible solutions. Our main theoretical contribution is an instance-dependent approximation ratio derived from the optimal Lagrangian dual multipliers, which strengthens the classical approximation bound of Nesterov, Roos, and Terlaky (NRT). The approximation is further refined using the Karush–Kuhn–Tucker (KKT) conditions and serves as a high-quality incumbent for an exact branch-and-bound algorithm. We further introduce a strategy for eigenvector branching that partitions branching intervals into multiple subintervals rather than relying on conventional bisection, substantially reducing the search tree. Computational results show that the KKT-refined solutions are globally optimal on the tested instances and that the proposed framework significantly reduces computation time compared with benchmark solvers. These results suggest that, for HQCQP, investing additional effort in solving a stronger Lagrangian relaxation can simultaneously improve approximation quality and accelerate exact global optimization.

Bicriteria Approximation Algorithms for Demand Matching

Yuchong Pan, MIT

The demand matching problem is a common generalization of the knapsack problem and the $b$-matching problem. In this problem, each edge of a graph has a demand and a weight, each vertex has a capacity, and the goal is to find a maximum weight subset of edges whose total incident demand at every vertex does not exceed its capacity. We study $(\alpha, \beta)$-bicriteria approximation algorithms for the demand matching problem, where the algorithm may violate each vertex capacity by at most $\beta$ times the maximum demand while outputting a solution whose weight is at least $1/\alpha$ times the optimum of the original instance. We present an iterative relaxation algorithm for the demand matching problem that exploits a structural characterization of strictly fractional extreme points of the natural LP relaxation. Combining this structure with a better-of-the-two rounding strategy yields $(7/6, 1)$- and $(1, 1)$-bicriteria approximation algorithms for general and bipartite graphs, respectively. We further generalize this approach to obtain a family of parametric bicriteria approximation algorithms, including a $(1, 4/3)$-bicriteria approximation algorithm. We also present a greedy, combinatorial $(k, 1)$-bicriteria approximation algorithm for the more general $k$-hypergraph demand matching problem, which includes the demand matching problem when $k = 2$. We complement these algorithmic results with lower bounds on the best possible weight approximation ratio with respect to the natural LP relaxation as a function of the additive capacity violation parameter $\beta \geq 0$. These lower bounds match the algorithmic guarantees for all $\beta \in \{ 0 \} \cup [1, \infty)$, thereby completely characterizing the trade-off between weight approximation and additive capacity violation in this regime. These bounds reveal a phase transition at $\beta = 1$: as $\beta$ approaches $1$ from below, triangle instances give a lower bound approaching $3/2$, whereas allowing one full unit of maximum-demand additive violation enables a $7/6$-approximation.

Averaged Quantile Randomized Kaczmarz Method for Noisy, Corrupted Linear Systems

Sofiia Shvaiko, Princeton University

Iterative methods for solving overdetermined linear systems are widely used in scientific computing and machine learning. However, their performance can be heavily affected in the presence of both measurement noise and corrupted observations, particularly when the two are difficult to distinguish. The Randomized Kaczmarz (RK) algorithm is a popular iterative solver that is capable of approximating the solution in the presence of noise, but is known to be fragile under large adversarial corruption. The recently introduced QuantileRK (QRK) algorithm, as well as its accelerated block version, addresses this challenge by employing a quantile threshold to identify untrustworthy rows. In this work, we present a convergence analysis of the averaged variant of QuantileRK. We establish convergence guarantees under appropriate separation conditions and support our analysis with numerical experiments investigating the combined effects of measurement noise and sparse corruption. Our results illustrate when separation between noisy and corrupted measurements is necessary for reliable convergence.

Prizes & Awards

We are pleased to announce significant increases to the competition budget for YinzOR 2026! A panel of three faculty judges will evaluate and select the top three presentations, and all attendees will vote for the "Fan Favorite" award.

Guidelines & Logistics

• Presentation Format: Selected presenters will have exactly 10 minutes to explain their research. Whiteboards or digital slides are permitted for visual support.
• Scoring: Presentations are scored based on clarity, impact, slide quality, and ability to communicate technical topics to a general OR/MS audience.
• Judges: Our faculty judges for YinzOR 2026 will be announced soon.