New York-Ontario (NYON) Operations Workshop

Program

7:45–8:30 - Breakfast

 

8:30–8:45 - Welcome Remarks | Ricky Roet-Green, University of Rochester

 

8:45–9:45 - AI / Innovation / Technology

Title: Dynamic Prediction of Waiting Time Intervals: Application to Cancer Care Facilities

Speaker: Yaron Shaposhnik, Simon Business School, University of Rochester. 

Coauthors: Yashi Huang (University of Rochester) and Arik Senderovich (York University). 

Abstract: We collaborate with a large cancer hospital and study the problem of predicting in real-time patients’ waiting time for examinations. We conduct an extensive set of numerical experiments using recent operational data, in which we compare a broad set of methodologies and report on their performance. We further develop and evaluate a prototype in a series of field experiments. We report on the models’ performance and discuss lessons learned from operationalizing and deploying the system. 

 

Title: From Descriptive to Predictive Analytics in Healthcare Service Operations: A Strategic Analysis of AI-Driven Digital Twins for Emergency Department Capacity Management 

Speaker: Opher Baron, Rotman School of Management, University of Toronto 

Coauthors: Shany Azaria (Tel Aviv University) and Dmitry Krass (University of Toronto) 

Abstract: Many modern business processes operate under congestion: stochastic demand, shared and limited resources, complex routing, and tight service-level constraints. From healthcare and public services to logistics, financial operations, and large-scale customer support, these systems exhibit nonlinear behavior where small disruptions propagate quickly and performance deteriorates sharply. This paper examines how Business Process Management (BPM) can evolve to address congestion-driven environments by integrating process mining, queueing theory, simulation, and machine learning and artificial intelligence (ML&AI) into digital twins. We discuss how descriptive analytics must move beyond static dashboards toward data-driven process models that reveal how processes actually unfold in practice—capturing routing variability, rework loops, synchronization delays, and resource contention. Process mining provides structural visibility; queueing-aware modeling explains performance; ML&AI supports prediction under uncertainty. Building on these foundations, predictive and comparative analytics enable counterfactual “what-if” evaluation of staffing, routing, prioritization, and scheduling policies before implementation. Finally, prescriptive analytics embedded within real-time digital twins allow organizations to intervene proactively—anticipating congestion, reallocating capacity, and mitigating cascading delays. Drawing on industrial deployments through SiMLQ, We demonstrate how combining process intelligence with congestion-aware operational modeling transforms BPM from retrospective analysis into continuous, real-time decision support. The implications extend across service industries where variability and resource contention define performance, resilience, and competitiveness. Emergency departments are not failing because clinicians lack commitment; they are failing because the operating system that coordinates demand, staffing, beds, diagnostics, specialist consultation, and inpatient transfers remains largely reactive. Digital twins that are based on extensive descriptive analytics can transform fragmented event data into a safe, continuously updated environment for prediction, counterfactual evaluation, and prescriptive action. The managerial implication is blunt: modern businesses, such as hospitals that rely only on after-the-fact statistics will remain trapped in decision blind spots. Businesses that build validated digital twins would move capacity management upstream, detect failure before it occurs, and redirect scarce resources to the highest-leverage bottlenecks in real time.

9:45–10:15 - Coffee Break

 

10:15–11:15 - Healthcare

Session Chair: Vera Tilson, University of Rochester

 

Title: The Impact of Stress on IVF Outcomes: A Predictive Framework 

Speaker: Vedat Verter, Queen's University 

Abstract: Physiological markers of stress have not been systematically evaluated as predictors of pregnancy success during in vitro fertilization (IVF). We investigated whether central, autonomic, and cardiovascular stress-related biomarkers measured non-invasively during IVF treatment are associated with clinical pregnancy. In this prospective observational study, women undergoing their first IVF cycle recorded daily morning measurements using a wearable electro-physiological device during a 7-day baseline period and throughout ovarian stimulation. The device quantified 35 autonomic and cardiovascular parameters reflecting sympathetic–parasympathetic balance, resilience, and homeostatic regulation. Ninety-four participants with ≥2 recordings in each period were included (clinical pregnancy rate: 54.3%). Supervised classification models were developed to predict clinical pregnancy. All possible three-feature combinations derived from 105 engineered features were evaluated. Logistic regression demonstrated the best performance, closely followed by ridge classification (alpha=1.15). Model performance was assessed using repeated stratified 80/20 train–test splits (1,000 iterations) to ensure stability. Statistical significance of selected predictors was evaluated via non-parametric bootstrapping (1,000 resamples). The best performing model achieved a mean predictive accuracy of 69% on unseen data, representing a 36% relative improvement over chance-level classification. Three autonomic features were independently associated with clinical pregnancy (bootstrap p<0.05). Higher aerobic readiness and energy supply during ovarian stimulation were positively associated with pregnancy, whereas increased cardiac tension during stimulation and elevated baseline low-frequency heart rate variability spectral power demonstrated negative associations. Larger studies with external validation are warranted to assess clinical utility and integration with established reproductive predictors.

 

Title: Epidemics in Queues 

Speaker: Jamol Pender, Cornell University 

Abstract: In this talk, we introduce and analyze a stochastic model for the spread of infectious diseases in service systems. The model is based on an M/M/1 queue in which a proportion p of arriving customers are infectious, while the remaining customers are susceptible. Transmission occurs through overlap in the queue: if an infectious and a susceptible customer are simultaneously present in the system for more than ℓ units of time, the susceptible individual becomes infected. We analyze the model from two complementary perspectives. From an individual (personalized) perspective, we study the infection risk faced by a single susceptible customer. In particular, we derive the probability that the customer becomes infected during their time in the system, as well as the probability that they are infected by at least k distinct infectious individuals. From a community perspective, we examine the spreading potential of a single infectious customer. Specifically, we compute the mean and higher moments of the number of susceptible customers infected by one infectious individual while they remain in the queue. Together, these results provide new insight into how transient interactions in service systems contribute to infectious disease transmission.

11:15–12:15 - Student Paper Competition: Finalists Presentations

 

12:15–1:15 - Lunch

 

12:45–1:15 - Student Poster Session

 

1:15–2:15 - Keynote Address

The Physical Backbone of AI: Supply Chains, Data Centers, and the Electricity Grid 

Plenary speaker: Serguei Netessine, Warton School 

 

Artificial intelligence is commonly described as a digital or software revolution. Yet every AI model rests on an extensive physical backbone: critical minerals, advanced semiconductor equipment, highly concentrated chip-fabrication capacity, servers and networking hardware, data centers, cooling systems, and enormous supplies of electricity. The ability to scale AI, therefore, depends not only on algorithms and data but also on the capacity, resilience, and coordination of these underlying supply chains. This talk follows the physical journey through which raw materials and energy are converted into computational intelligence. Many stages of this journey exhibit extreme supplier concentration, long capacity-expansion lead times, limited substitutability, and rapidly growing but highly uncertain demand. These characteristics create bottlenecks, geopolitical vulnerabilities, and the potential for classic supplychain dynamics such as overordering, shortages, and costly excess capacity. Electricity provides the most immediate illustration of these challenges. AI data centers create large, geographically concentrated, and nearly continuous loads that can exceed the capabilities of local generation and network infrastructure. Using facility-level data on data-center capacity and a quarterly utility panel covering the United States, we discuss that data-center expansion increases retail electricity tariffs, but the effects vary substantially across utilities and customer groups. Utilities with prior experience serving large industrial customers are better positioned to accommodate this demand without passing substantial costs on to households. The emerging AI race is therefore also an infrastructure and operations race. Its outcome will depend on whether firms and governments can coordinate investments across semiconductor supply chains, datacenter capacity, electricity generation, transmission, and operational flexibility. The talk concludes by examining how improved siting, rate design, supply-chain resilience, and more flexible computing and energy systems can support AI growth without allowing physical-world bottlenecks to constrain progress in the digital one.

2:15–3:15 - Service Operations

Session Chair: Burak Kazaz, Syracuse University

 

Title: Dynamic Optimization of Workforce Talent 

Speaker: Parshan Pakiman, University at Buffalo 

Abstract: We study the problem of repeatedly assigning recurring jobs to a pool of workers under two dynamics: (i) workers gain or lose familiarity with each job type over time based on whether or not they are assigned to the job, and (ii) the availability of workers and jobs evolves stochastically. Learning from experience incentivizes building specialized workers, whereas uncertain availability incentivizes maintaining cross-functionality. We formalize this key management trade-off as a Markov decision process with endogenously evolving worker-job familiarity levels and combinatorial constraints that model feasible assignments under job and worker availability. We develop two complementary classes of assignment policies. The familiarity-agnostic (FA) policy prescribes assignments based on availability, abstracting away from current familiarity levels. It approximates the value of specialization through limiting familiarity levels while capturing cross-functionality by varying assignments across availability realizations. We establish a performance bound showing that the FA policy is near-optimal when learning and forgetting are slow or job arrivals are sparse, and derive a quadratic program to optimize it. The Lagrangian relaxation (LR) approach relaxes the combinatorial constraints and penalizes violations using dual variables. It captures specialization through pairwise familiarity dynamics and approximates cross-functionality through dual penalties. The resulting relaxed policy has a familiarity-threshold structure that favors specialization. We recover a feasible LR policy through projection and randomization, establish a performance bound, and prove optimality under deterministic availability. Numerical experiments show that the proposed policies outperform or match competitive benchmarks while remaining computationally practical and reveal when specialization and cross-functionality are most valuable.

 

Title: Conservative Dynamic Pricing with Demand Learning 

Speaker: Yun Zhou, McMaster University 

Abstract: Algorithms for dynamic pricing with learning can suffer early revenue losses leading managers to abandon them in favor of baseline policies. We propose a pricing algorithm that provides safety guarantees by ensuring with high probability that revenue achieves at least a specified fraction of a baseline policy’s performance. We first study stage-wise safety constraints which require expected revenue in each period to meet a minimum benchmark. We develop a UCB-based algorithm and show that it achieves near- optimal learning performance while maintaining these guarantees. We then consider cumulative safety constraints on total revenue and establish similar performance bounds. Additional extensions examine checkpoint-based safety evaluation and a unified framework that jointly handles price optimization and constraint enforcement. Title: Conservative Dynamic Pricing with Demand Learning

3:15–3:45 - Coffee Break

 

3:45–4:45 - Supply Chain Analytics

Session Chair: Milind Sohoni, University at Buffalo

 

Title: Human-Centric Perishable Inventory Management with AI Assistance 

Speaker: Meng Qi, Cornell University 

Coauthors: Yu Nu, Elena Belavina, Karan Girotra 

Abstract: Improving production and inventory decisions in food service systems is challenging due to product perishability, demand censoring, and the fact that inventory decisions are made by human decision makers subject to biases and misaligned incentives. These features limit the applicability of existing inventory control methods and complicate the deployment of algorithmic decisions in practice. We develop a data-driven approach that learns the optimal base-stock policy directly from historical sales data while explicitly accounting for both product perishability and demand censoring. The resulting inventory policy is supported by theoretical guarantees that include asymptotic consistency, asymptotic normality, and bounds on the long-run cost gap. We operationalize the learned policy as a prescriptive assistant that recommends inventory decisions to human managers. We further propose and compare interventions that increase compliance with our recommendations, including performance feedback and uncertainty-aware prescriptions that enhance algorithm transparency. Going beyond existing interventions that are primarily algorithm-focused, we adopt a new human-centric perspective through a bias detection system that flags deviations from optimal actions and diagnoses the behavioral biases that may have caused them. It serves not only as a compliance intervention but as an alternative form of algorithmic decision support, which we term a detective assistant. In an online experiment with Prolific workers, both assistants improve human decision quality when deployed individually. Relative to a human-only baseline, the prescriptive assistant reduces total inventory-management costs by about 25%, while the biasdetection assistant reduces costs by approximately 19%. When used to augment the prescriptive assistant, both bias detection and confidence-interval-based uncertainty communication reduce costs by approximately 41% and 40%, respectively, substantially outperforming standard compliance interventions.

 

Title: Do Penalties Promote Responsible Operations? Evidence from the U.S. Mining Industry 

Speaker: Vibhuti Dhingra, York University 

Abstract: Financial penalties---fines levied for violations found during site inspections---are widely used to encourage responsible operations, such as compliance with labor and environmental standards or with supplier codes of conduct. However, the effectiveness of penalties in achieving this objective remains unclear. In this paper, we empirically investigate whether higher penalties lead to more responsible operations with a focus on workplace safety. Using a large, granular dataset from the U.S. mining industry, our identification strategy exploits a policy shock that sharply increased the penalty amounts for health and safety infractions in the mines. We find that higher penalties had, at best, a modest effect on responsible operations. While there was some decrease in violations, which was heterogeneous across mine types, we find no evidence of a decrease in accidents or injuries at mines that faced higher fines. We identify several mechanisms driving this null effect. First, the violations that decreased were unrelated to the root causes of accidents. Second, operational complexity does not explain the heterogeneity across mines, but labor market power does: penalties were least effective in counties with high unemployment where workers have fewer outside options. Third, the decline in violations was concentrated at mines that were already relatively safe before the penalty increases. Finally, we identify an unanticipated and unintended consequence of higher penalties that compromised their effectiveness: mining companies strategically contested the higher fines in court, which led to discounted payments, delay in payment times, and an overflow of the court's operating capacity. By examining the mechanisms through which higher penalties (fail to) promote responsible operations, our work offers insights that can be generalized to contexts beyond mining (e.g., to global supply chains).

4:45–5:00 Closing Remarks & Announcements | Ricky Roet-Green, University of Rochester

 

5:00–6:00 - Reception