Data-Driven Optimization under Uncertainty: From Modeling to Scalable Algorithms
Seminar
Speaker
Shimrit Shtern (Technion)
Date
21/12/2025 - 13:30 - 12:00Add to Calendar
2025-12-21 12:00:00
2025-12-21 13:30:00
Data-Driven Optimization under Uncertainty: From Modeling to Scalable Algorithms
Abstract: Many decision-making problems under uncertainty can be formulated as two-stage stochastic optimization problems. In such problems, a decision maker must choose an initial (first-stage) decision without knowing the realization of the uncertainty, followed by a recourse (second-stage) decision once the uncertainty is revealed. The first-stage decision must therefore minimize the expected value of a cost function, and together with the recourse decision, satisfy problem-specific constraints for all possible realizations of the uncertainty. A key challenge stems from the fact that the true probability distribution of the uncertainty is typically unknown, and only historical data is available. A common approach, Sample Average Approximation (SAA), directly optimizes over this data. However, this can lead to overfitting, resulting in poor out-of-sample performance or even infeasible solutions.In this talk, I present a practical and scalable framework for handling uncertainty in large-scale, data-rich environments—from modeling to solution algorithms. At the core of this approach is the Sample Robust Optimization (SRO) paradigm, which mitigates overfitting by optimizing against worst-case perturbations around each data point. Although solving SRO exactly is computationally hard, we show that it can be effectively approximated using local-linear decision rules, with asymptotic optimality guarantees as the number of samples increases. Crucially, the resulting models are compatible with standard first-order optimization methods such as stochastic gradient descent, which enables efficient implementation even with large data sets. We demonstrate the practical benefits of our approach through applications in multi-location transshipment and portfolio optimization, highlighting its potential for real-world decision-making across diverse domains.
Math seminar room (201), Math building (216)
אוניברסיטת בר-אילן - המחלקה למתמטיקה
mathoffice@math.biu.ac.il
Asia/Jerusalem
public
Place
Math seminar room (201), Math building (216)
Abstract
Abstract: Many decision-making problems under uncertainty can be formulated as two-stage stochastic optimization problems. In such problems, a decision maker must choose an initial (first-stage) decision without knowing the realization of the uncertainty, followed by a recourse (second-stage) decision once the uncertainty is revealed. The first-stage decision must therefore minimize the expected value of a cost function, and together with the recourse decision, satisfy problem-specific constraints for all possible realizations of the uncertainty. A key challenge stems from the fact that the true probability distribution of the uncertainty is typically unknown, and only historical data is available. A common approach, Sample Average Approximation (SAA), directly optimizes over this data. However, this can lead to overfitting, resulting in poor out-of-sample performance or even infeasible solutions.
In this talk, I present a practical and scalable framework for handling uncertainty in large-scale, data-rich environments—from modeling to solution algorithms. At the core of this approach is the Sample Robust Optimization (SRO) paradigm, which mitigates overfitting by optimizing against worst-case perturbations around each data point. Although solving SRO exactly is computationally hard, we show that it can be effectively approximated using local-linear decision rules, with asymptotic optimality guarantees as the number of samples increases. Crucially, the resulting models are compatible with standard first-order optimization methods such as stochastic gradient descent, which enables efficient implementation even with large data sets. We demonstrate the practical benefits of our approach through applications in multi-location transshipment and portfolio optimization, highlighting its potential for real-world decision-making across diverse domains.
In this talk, I present a practical and scalable framework for handling uncertainty in large-scale, data-rich environments—from modeling to solution algorithms. At the core of this approach is the Sample Robust Optimization (SRO) paradigm, which mitigates overfitting by optimizing against worst-case perturbations around each data point. Although solving SRO exactly is computationally hard, we show that it can be effectively approximated using local-linear decision rules, with asymptotic optimality guarantees as the number of samples increases. Crucially, the resulting models are compatible with standard first-order optimization methods such as stochastic gradient descent, which enables efficient implementation even with large data sets. We demonstrate the practical benefits of our approach through applications in multi-location transshipment and portfolio optimization, highlighting its potential for real-world decision-making across diverse domains.
תאריך עדכון אחרון : 18/12/2025
