Generative Inverse Optimization
Generative AI has swept the nation, in many ways leaving traditional optimization behind. However, it ironically borrows heavily from the optimization toolbox.
In this project, we look to generate models from large datasets, in a similar fashion to AI. Unlike AI, however, we will be working with datasets that form optimization problems and learning those instead.
Through this project, we hope to learn more about the overlap between modern-day AI and OR, and how the two fields can continue to build each other up over time.
Learning to Pivot (SimplexRL)
The simplex method's speed depends on which pivot it takes at each step, and classical rules choose by fixed local formulas. In this project, led by Felix Parker with Emmett Springer and Kimia Ghobadi, we model pivot selection as a sequential decision problem, use exact search to find the shortest routes, and train policies to follow them.
An early learned policy already ties steepest edge on a 40 x 40 test problem. A companion paper on learned warm starts, with me as co-first author, is on arXiv and under review at ICLR 2027. See the pivot timeline in action in my Walking the Corners lecture.
The Impact of AI Scribes on Primary Care Practitioners
AI scribe technology has been rapidly deployed in practitioners' offices in the past year or two, but the studies in terms of efficacy and where it is and isn't helpful are just catching up. We are working on one of these studies!
The Homecare Provider Allocation Problem
Dispatching home care workers to patients is a natural problem for operations research, in that it is difficult, multistage and data-intensive. In this project, we look for ways to make this very large problem tractable and adjustable for non-expert users that will be using the end-systems.
Incorporating Robustness into Radiation Therapy Treatment Planning
When planning radiation therapy plans, or optimizing in general, the greediness of the underlying algorithms can have negative impacts on plan selection. In this multi-year body of work, we investigate the interplay between very large models and uncertainty, studying which kinds of approaches are best to mitigate uncertainty in various scenarios.
Geometric Beam Angle Optimization
When clinicians plan radiation therapy treatment plans for step-and-shoot radiation therapy, the beam angles must be chosen before any calculation can be made. These angles are often selected to be equidistant, due to the difficulty in determining the best angles a priori. The literature has employed heuristic approaches that have been shown to work better than this naive equidistant approach. However, what if there were more knowledge-based approaches to doing so?
Radiation Therapy Sampling and Solution Pipeline/Add-On
A MATLAB-based codebase that takes in matRad-calculated dose objects, allows for multiple sampling modalities, and then connects to a CPLEX-c++ solver for optimization.
We used it to compare eight voxel sampling approaches on clinical cases, published in Medical Physics (2025).
Light Pareto Robust Optimization
Robust optimization protects radiation therapy plans against breathing motion, but it plans for the worst case and pays for it in every other scenario. Light Pareto robust optimization keeps the plan robust on the tumor, then allows a small, controlled amount of extra heart dose in exchange for much better plans in typical breathing scenarios. On five left-sided breast cancer patients, each with 50 simulated breathing patterns, plans began to look deliverable with as little as 0.1 Gy of extra heart dose on average, and most patients saw diminishing returns beyond 0.2 Gy. In the figure, the robust (RO) and Pareto robust (PRO) plans overdose the target, while light Pareto robust plans fall off sharply at the prescribed dose.
Published in Medical Physics (2023), with F. Rahimi, H. Abouee-Mehrizi and H. Mahmoudzadeh.
Robust Direct Aperture Optimization
Radiation therapy plans are usually built in stages: doses first, then the aperture shapes the machine delivers, which can cost plan quality. Robust direct aperture optimization does both at once and protects the plan against breathing motion, in a single mixed-integer model that includes the machine's deliverability rules. To make it solvable, I designed a candidate plan generation heuristic that quickly produces clinically viable plans and warm-starts the full model.
Published in the INFORMS Journal on Computing (2022), with T. G. Purdie, T. C. Y. Chan and H. Mahmoudzadeh.
morDiRECT: An interface for comparing tradeoffs in radiation therapy treatment planning
This paper took a preliminary look at how to handle black-box multi-objective problems, with the specific use case of stereotactic radiosurgery planning.
Conceptually, the interface took the idea of multi-objective trade-offs and used this as an engine for delivering plan alternatives to the end user. The interface facilitated in-depth trade-off analysis between a set of plans that were prerun, for greater flexibility and an intuitive understanding of the decision space.
Published in Expert Systems with Applications (2015), with D. M. Aleman and K. Ghobadi.
Simple PLINKO Game
Inspired by the Price Is Right minigame, I used this simple, visual R program to simulate one or multiple games of PLINKO for statistics students at the University of Waterloo.
Since Markov chains are slightly out of scope of the course, we focus on the binomial distribution and the similarities and differences that come from that. Note, this application allows for non-standard board generations, which makes this teaching process easier. Message me to connect if you would like a copy of the labs!
SuperSlider
A flexible MATLAB-based slider creation tool that allows for multiple slides along a single axis. This was built out for an in-house interface project, but has been downloaded by over 1000 people and put into use in some cool industries!
If you want your users to control a multi-point range, give it a download and let me know what you're working on!
IIE Arena Simulation Competition
Back in 2013, Krystle Pang, Jaquelyn Monis Rodriguez and I set out to use Arena Simulation Software to predict the impact of a new terminal rollout at Chicago's O'Hare International Airport. I believe we came in 3rd at the International competition.
Recycling Sorter
Back in 2012, I was in Mechanical Engineering and Yuhin Wu, Ryan Brown, Keith Tong and I spent many a late night at Skule™ building ART: the automated recycling terminal. ART was an Arduino-based recycling sorter for ASME's SEC competition (and we won the local competition!). We were all very young here.
If you pay attention to the bottles in the videos, you might notice they break most conventional sorting mechanisms -- this was not an accident! Can you tell how the machine works?
Learning Clinical Models from Archives
Hospitals hold years of treatment plans, schedules and menus, many with expert ratings attached. I will scale generative inverse optimization to these archives, starting with open radiation therapy data from the PortPy library, so that the model behind expert decisions can be learned, read and corrected by clinicians, and I will establish when a learned model can be trusted for a new patient.
AI and OR are governed by the same mathematical principles, so the reverse direction matters too: the same tools could make learned AI systems explainable, and help them run more efficiently.
Solvers that Learn from Search
Large healthcare models are solved again and again, while classical solvers follow the same fixed rules every time. Search can find far shorter paths than any fixed rule, and learned policies can follow them. I will extend learned pivots and warm starts to the mixed-integer and decomposition methods that healthcare models depend on, measure the gains in run time as well as iterations, and release them in open-source solvers such as HiGHS, so clinics without commercial licenses benefit too.
New Formulations for Direct Aperture Optimization
Direct aperture optimization designs the exact shapes a treatment machine delivers, but it is hard to solve while also limiting how many apertures a plan uses, a limit that both the machines and plan quality demand. Building on joint work with Isaac Rudich (Polytechnique Montréal), I am developing a network-based formulation built around shortest paths, with constraint-and-column generation and heuristics that rule out shapes the machine cannot deliver. I am also pursuing a new branch-and-bound approach.
Once these solve the basic cases quickly, the next step is adding the robust and CVaR dose constraints that clinics need.
Bridging AI Progress and Clinical Trust
AI models are improving faster than clinicians can learn to trust them. I will work to close that gap through transparency and explainability, and by measuring where AI improves care operations and where it gets in the way. I plan to build these models directly from electronic health record data, so they reflect real clinical workflows and can be evaluated where clinicians use them.
Novel Mixed-Modality Imaging Techniques
With the current advent of AI and advanced OR techniques, I believe humanity is about to become a lot more transparent. I am interested in studying less invasive, more informative imaging protocols that we have the technology for now, not decades from now. Please reach out if you would like to collaborate or to know more!