Timothy C. Y. Chan

When “Optimal” Advice Gets Ignored Algorithms can recommend decisions that look optimal on paper but conflict with how people assess the situation, making those recommendations less likely to be followed. In “Conformal Inverse Optimization for Adherence-Aware Prescriptive Analytics,” Chan, Delage, and Lin address this challenge by using past decisions to learn how people evaluate different options. Because those decisions may reflect diverse preferences, noisy observations, or imperfect models, the authors do not rely on a single estimated preference profile. Instead, they construct an uncertainty set around the inferred parameters and use robust optimization to generate recommendations that perform well both objectively and from the decision maker’s perspective. Their method comes with statistical and performance guarantees and outperforms standard inverse optimization in numerical experiments. In a simulated Toronto food-delivery study, it substantially improves estimated courier adherence, maintaining comparable delivery times, showing how better-aligned recommendations can benefit both users and organizations.

The amount and quality of active transportation infrastructure directly influence how much people cycle. However, we still lack quantitative tools to predict shifts in trips and mode choice as a result of investments in cycling infrastructure, which results in opinion- rather than evidence-based debates about the value of infrastructure. In this work, we model bicycle route choice, a key element in modeling cycling demand. Unlike motor vehicles which are sensitive mainly to distance and travel time, cyclists frequently detour from the shortest path to access safer and more pleasant routes. We combine over 3000 GPS cycling trajectories from the Toronto Household Activity-Travel Survey with road network, cycling network, and land use attributes to estimate bicycle route choice parameters. We systematically evaluate the impact of several modeling thresholds that have been arbitrarily chosen in previous works and show that results are more stable at large choice set sizes and when circuitous trips are removed. We quantify cycling infrastructure and road geometry with Level of Traffic Stress (LTS), a widely-used rating system for cycling comfort. LTS is correlated with cycling mode choice and segment-level cycling counts but has rarely been applied in bicycle route choice modeling. We estimate route choice parameters using a discrete choice modeling framework and compare two formulations of the model using either LTS or its underlying road class features to evaluate LTS as a route choice predictor. We find consistent patterns between LTS and road class, finding in the LTS model that cyclists prefer LTS 1 and LTS 2 and avoid LTS 4 compared to LTS 3, and finding in the road class model that cyclists prefer infrastructure relative to no infrastructure and prefer medium and local roads over major roads.

Cycling is an effective way to promote healthy lifestyles, improve urban mobility, and combat climate change, but safety and comfort concerns remain major barriers to its global uptake. To make cycling safer and more inviting, we collaborated with the City of Toronto to develop and implement a suite of analytics tools to assess and plan urban bike infrastructure. This collaboration demonstrates how operations research tackles key transportation planning challenges with tangible real-world impact, and leads to methodological development that contribute to advancements in machine learning and optimization.

Scarce data hinders data-driven location optimization, as sample average approximation (SAA) can yield poor out-of-sample performance. While target data is limited in these settings, abundant related data may be available as a proxy. We study mixed-effects (ME) policies, which combine unbiased, noisy historical-data-based estimates with biased, low-variance proxy estimates, for facility location problems with small data. We analyze a stylized three-node example to derive insights into when ME policies outperform the gold policy (i.e., SAA using historical data) and extend our analysis to general location problems based on a softmax approximation. Our results show that ME policies are most effective in low signal-to-noise ratio domains where the proxy is sufficiently aligned with the target variable. Notably, even imperfect or seemingly low-quality proxies can be leveraged in ME policies to outperform the gold policy in both expected and lower-tail performance. A case study on locating critical care response units for response to major trauma incidents demonstrates that ME policies based on lower-acuity trauma as a proxy for major trauma outperform the gold policy in expected coverage and solution stability. Our work demonstrates that out-of-sample performance of location decisions in the small data regime can be strengthened by incorporating proxy information within the optimization framework. We provide analytical conditions and practical guidance on when proxies are valuable, and demonstrate through a case study that ME policies are straightforward to implement to improve expected performance and policy robustness.

Adaptive robust optimization problems have received significant attention in recent years but remain notoriously difficult to solve when recourse decisions are discrete. In this paper, we propose new reformulation techniques for adaptive robust binary optimization (ARBO) problems with objective uncertainty. Without loss of generality, we focus on ARBO problems with “selective adaptability,” a term we coin to describe a common class of linking constraints between first-stage and second-stage solutions. Our main contributions involve reformulating and approximating these ARBO problems as network flow models by leveraging ideas from the decision diagram literature. We show that these models are versatile and easy to customize and can generate feasible solutions, primal bounds, and dual bounds. Furthermore, in contrast to existing approaches, they require no specialized algorithms and can be solved directly using standard optimization solvers. Through an extensive set of computational experiments, we show that our models generate high-quality solutions and dual bounds in significantly less time than popular benchmark methods. History: Accepted by Andrea Lodi/Design & Analysis of Algorithms-Discrete. Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information ( https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2024.0718 ) as well as from the IJOC GitHub software repository ( https://github.com/INFORMSJoC/2024.0718 ). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/ .

“Superstar” markets, where a small portion of individuals earn disproportionately high incomes, are common in fields like entrepreneurship, sports, and entertainment. Participants in these markets face significant income uncertainty, which can deter entry or prompt early exit. To address this difficulty, we propose income pools: contracts where individuals agree to share a portion of future earnings with pool members if a specific salary milestone is achieved. To date, hundreds of income pool contracts have been signed in practice. This paper develops the first mathematical model to analyze such contracts, focusing on stability (i.e., pools where no agents leave or join). When constrained to creating a single pool, we show that a stable pool always exists with specific structural properties. Generally, stronger agents tend to be more “collaborative” and favor larger pools, whereas weaker agents are more “selfish” and prefer smaller pools. Next, we analyze pools that adhere to a maximum size (mirroring current practice) and show that stability persists. These pools typically require more weaker agents than strong ones and we find an interesting “Pareto dominance” result, whereby all agents in a stable pool prefer a particular unique stable pool. Finally, we study general partition structures, prove that a stable partition always exists under certain conditions, and provide an algorithm to construct such a partition. We conclude with a case study on data from 2,000 professional baseball players to demonstrate a 20%–30% increase in social welfare if players join income pools, under varying contract parameters. This paper was accepted by Victor Martinez de Albeniz, operations management. Funding: T. C. Y. Chan received funding support from the Natural Sciences and Engineering Research Council of Canada. N. Chen has no funding sources to report. C. Fernandes received funding support from the Natural Sciences and Engineering Research Council of Canada and the TD Management Data and Analytics Lab. Supplemental Material: The online appendix and data files are available at https://doi.org/10.1287/mnsc.2024.07480 .

Background and aims: The macronutrient composition of donor human milk (DHM) can vary substantially due to several factors such as maternal age, diet, and lactation duration. However, consistent macronutrient levels in DHM facilitate the administration of the required amounts to preterm infants. The current pooling strategy at most human milk banks combines milk from different batches from a single donor. This study aims to stabilize the macronutrient quality of DHM by pooling milk from different donors by utilizing machine learning prediction and optimisation techniques. Methods: The current pooling strategy is compared with a new theoretical approach that pools milk batches from up to 5 donors. To predict the crude protein and energy content, we used the following variables: body mass index, the donor’s diet (vegetarian or non-vegetarian), maternal age, full-term or preterm delivery, lactation stage, and volume pumped. These predictions are then used within an optimisation model to create milk pools that minimize the deviations from the target macronutrient levels (1.0 g protein/100 mL and 70 kcal/100 mL). Results: The prediction model is based on 2236 created single-donor pools from 480 donors. Random forest regression models provided the most accurate predictions of macronutrient content. The new pooling strategy using multiple donors shows reduced deviations from target values compared to the current single-donor approach (average total absolute deviation 0.402 versus 0.664). Conclusion: This study proves the potential of data-driven methods to improve operational efficiency in human milk banks, and improving the consistency of donor human milk.

Here you can find the data (data folder), code (src folder) and main results (results folder) from the numerical experiments found in our paper. The data corresponds to the particular model parameterizations of each adaptive robust problem that we considered. The code is written in Python, and covers the process of building a decision diagram, as well as formulating and solving the corresponding network flow models using Gurobi. Finally, the results include the solution times, primal and dual bounds generated by our models. Additional details can be found in the README files within each folder.

Problem definition: Geographical mismatch between demand for care and healthcare resources is a critical challenge for hospital networks, particularly during episodes of surge such as pandemics, mass casualty events, and natural disasters. In such contexts, interhospital patient transfers can play a vital role in balancing congestion and promoting equitable access to care. We propose and investigate interhospital patient transfer policies to alleviate hospital congestion and inequity in the distribution of patients across the health system. Methodology/results: We propose a queueing network model capturing salient features of patient flow between acute wards and intensive care units (ICUs) both within and across hospitals. We formulate a stochastic control problem to determine optimal transfer policies and develop a solution method by leveraging a deterministic fluid approximation. Using the COVID-19 pandemic as a case study, we estimate and validate our model with data from 17 hospitals in Ontario, Canada. Following our policy, the expected number of patient-days above 95% capacity can be reduced by up to 48.4% in wards and 26.4% in ICUs, improving inequity in COVID patient distribution by up to 31.1% compared with the no-transfer policy. Our approach also outperforms historical transfer decisions that focused primarily on addressing COVID load inequity. We show that minimizing COVID load inequity alone can exacerbate existing imbalances in overoccupancy. In contrast, jointly optimizing both objectives leads to a very small compromise in COVID load inequity for a significant improvement in overoccupancy costs. These benefits are achievable by transferring only COVID patients and exhibit diminishing returns as the number of transfers increases. Managerial implications: Guiding patient transfers using our approach can effectively address imbalanced congestion and inequity during episodes of surge. These benefits can be obtained with relatively few transfers. Funding: The development of the GEMINI data platform used in this study has been supported with funding from the Canadian Cancer Society, the Canadian Frailty Network, the Canadian Institutes of Health Research, the Canadian Medical Protective Association, Green Shield Canada Foundation, the Natural Sciences and Engineering Research Council of Canada, Ontario Health, the St. Michael’s Hospital Association Innovation Fund, and the University of Toronto Department of Medicine with in-kind support from partner hospitals and the Vector Institute. We also acknowledge the support of the Government of Canada’s New Frontiers in Research Fund [Grant NFRFR-2022-00209]. Supplemental Material: The online appendix is available at https://doi.org/10.1287/msom.2025.0497 .