4 Presentations

• 
  02:45 PM - 03:07 PM

  Distributionally Robust Universal Classification

  • Weijun Xie, presenter, Georgia Institute of Technology
  • Siyuan Chen, Georgia Tech

  We study a distributionally robust optimization (DRO) formulation of a stochastic multi-class classification problem (SP) under the Wasserstein distance. The DRO formulation addresses the poor out-of-sample performance and overfitting issues of the sample average approximation (SAA) approach, which relies on the empirical distribution rather than an uncertainty set. To manage the infinite-dimensional nature of the DRO formulation, we develop an in-sample counterpart, referred to as in-sample DRO, which enables a tractable solution while maintaining robust properties. We establish asymptotic relationships between the original SP, the DRO formulation, and the in-sample DRO formulation, and we propose sample size guarantees to ensure reliable performance. Additionally, we derive a mixed-integer linear programming (MILP) representation to solve the in-sample DRO formulation efficiently, mitigating the curse of dimensionality over feature space. Numerical experiments highlight the improved effectiveness and resilience of the distributionally robust model compared to traditional classification techniques.

• 
  03:07 PM - 03:29 PM

  Machine Learning-Enabled Stochastic Capacity Expansion at Scale

  • Taehyeon Kwon, presenter, Pennsylvania State University
  • Anirudh Subramanyam, Pennsylvania State University

  We present a novel approach to addressing large-scale capacity expansion problems by integrating machine learning techniques with multi-horizon stochastic programming (MHSP). We propose a novel solution method that replaces the value functions of MHSP problems with machine learning surrogates. A scalable structure-exploiting sampling algorithm first trains the surrogate models, which are then embedded within mixed-integer optimization problems and solved with standard solvers. Our solution method is exemplified on EMPIRE, an extremely large capacity expansion planning model for Europe. The results show our approach achieves a notable reduction in solution time while also decreasing the variance of the objective value compared to existing methods. We believe our work can guide future studies on infrastructure development and resource allocation across the power and energy sectors.

• 
  03:29 PM - 03:51 PM

  Investigating the Monte-Carlo Tree Search Approach for the Job Shop Scheduling Problem

  • Laurie Boveroux, presenter, University of Liege
  • Damien Ernst, University of Liege
  • Quentin Louveaux , University of Liege

  The Job Shop Scheduling Problem (JSSP) is a well-known optimization problem in manufacturing, where the goal is to determine the optimal sequence of jobs across different machines to minimize a given objective. In this work, we focus on minimising the weighted sum of job completion times. We explore the potential of Monte Carlo Tree Search (MCTS), a heuristic-based reinforcement learning technique, to solve large-scale JSSPs, especially those with recirculation. We propose several Markov Decision Process (MDP) formulations to model the JSSP for the MCTS algorithm. In addition, we introduce a new synthetic benchmark derived from real manufacturing data, which captures the complexity of large, non-rectangular instances often encountered in practice. Our experimental results show that MCTS effectively produces good-quality solutions for large-scale JSSP instances, outperforming our constraint programming approach.

• 
  03:51 PM - 04:13 PM

  An imitation-based learning approach using DAgger for the Casual Employee Call Timing Problem

  • Antoine Legrain, GERAD - Polytechnique Montréal
  • Louis-Martin Rosseau, Polytechnique Montréal
  • Prakash Gawas, presenter, Polytechnique Montreal

  Predictive models are increasingly important in enhancing decision-making processes. This study proposes an innovative approach utilizing DAgger, an imitation learning algorithm, to iteratively train a policy for addressing stochastic sequential decision problems. These problems can be challenging, especially when expert input is costly or unavailable. Our focus lies in crafting an effective expert within the DAgger framework, drawing from deterministic solutions derived from contextual scenarios generated at each decision point. Subsequently, a predictive model is developed to mimic the expert’s behavior, aiding real-time decision-making. To illustrate the applicability of this methodology, we address a dynamic employee call-timing issue concerning the scheduling of casual personnel for on-call work shifts. The key decision involves determining the optimal time to contact the next employee in seniority order, allowing them to select a preferred shift. Uncertainty arises from the varying response times of employees. The goal is to strike a balance between minimizing schedule changes induced by early notifications or calls and avoiding unassigned shifts due to late notifications. Unlike traditional predict-and-optimize approaches, our method utilizes optimization to train learning models that establish connections between the system’s current state and the expert’s wait time. We apply our algorithm using data provided by our industrial partner to derive an operational policy. Results demonstrate the superiority of this policy over the current heuristic method in use.