I wrote a piece for ORMS Tomorrow called "PhD: My mistakes, my lucky breaks, and what I tell my students now". This article draws from my own experience as a student & from what I have learned by advising PhD students so far. Thank you Dan Ni Lin for the invitation! www.informs.org/content/down...

But if I were to stay here for longer than 4 hours, would Google Scholar notice it?

Our preprint will be out soon! 7/7

As a result, we explore solutions at the feasibility boundary, the results are consistently above the intended feasibility rate, and the runtime is considerably smaller than the standard approach. 6/N

D) We use cheaper alternatives, including an heuristic and a graph neural network, to make the feasibility check faster. 5/N

We address that as follows: A) We solve the optimization problem using a single scenario. B) To avoid overfitting, we multiply the right-hand side of the constraint by a parameter. C) We adjust the value of this parameter by cheaply using a massive set of scenarios to evaluate feasibility. 4/N

2) In the case of binary variables, most solutions satisfy either all or none of the constraints. Hence, training a machine learning model to predict feasibility requires careful curation of its training set. 3/N

Chance-constrained optimization has peculiar challenges: 1) If we optimize while ensuring feasibility, such as with sample average approximation, the feasibility gets slightly overfitted to the scenarios used. 2/N

Wrapping up #cors2026, Justin Dumouchelle talked about our joint work (along with Shunyu Yao and Beste Basciftci) on decoupling optimization from feasibility when solving chance-constrained optimization models. 1/N

He share interesting observations: - only delegating matrix multiplication to the GPU doesn’t yield a significant speedup - gaming GPUs are enough to obtain good results - an LP with more than 10M coefficients is solved faster in GPUs - there is a lot yet to be done for solving MILPs with GPUs 2/2

Carlos Zetina talked at #cors2026 about the benefits of GPU acceleration in solving linear programs using the FICO Xpress solver. Their solver has the two most popular methods implemented: - PDHG is faster in the problems that it can solve - cuPDLP+ solves more problems 1/2

One use case of machine learning in optimization is circumventing the cost of optimization algorithms. Jerry Sun presented such a case at #cors2026 by predicting Farkas multipliers to prove that a solution is infeasible with respect to the linear relaxation without solving the linear relaxation.

Existing approaches, such as Lagrangian relaxation and approximate dynamic programming, end up representing unreachable states. By augmenting the arcs labels with the history of nodes previously visited, she manages to improve performance with little impact on computational costs. 2/2

Seyedeh Parisa Moosavi talked at #cors2026 about relaxations for weakly-coupled Markov decision processes. She considered an example of managing fatigue in crew scheduling: teams need a break to recharge, but their breaks need to be coordinated with the breaks of other teams. 1/2

This Tuesday I am giving a tutorial on constraint learning at #cors2026 When: Tuesday, June 9, 10:45 AM Where: Walter Light Hall 205

We can reduce uncertainty in inverse learning by working with secondary objectives, so that we use expert knowledge to make some regions of the feasible set preferable. I was not able to capture the latter part of the talk, but you came read more here: arxiv.org/pdf/2603.17033 4/4

With inverse learning (if I got it right), we assume that the alternate (optimal solution of the forward optimization problem) is unique, but also that we do not know it with certainty. 3/N

My understanding of Ghobadi’s argument about identifiably in inverse optimization is that this problem naturally has multiple solutions, save some exceptions. We may deal with that by either change what we identify (inverse learning) or quantify what we can’t (MLE confidence regions). 2/N

Kimia Ghobadi talked at #cors2026 about inverse optimization challenges: - solution of the problem (the closest cost vector making a solution optimal) may not be unique - solving it becomes expensive with more observations - it is difficult to establish confidence on the retrieved parameters 1/N

Danielle Ripsman talked at #cors2026 about how to combine inverse optimization and generative models in healthcare applications, proposing a generative inverse optimization model.

In order to facilitate interpretability, Shiravi explained how to narrow down the set of generated constraints to a reasonable but representative size. 2/2

Rana Shiravi talks at #cors2026 about using inverse optimization to understand dosage tradeoffs in radiotherapy treatments. 1/2

Mahya Hemmati talks at #cors2026 about solving multistage stochastic optimization problems by approaching them as a two-stage rolling horizon stochastic problem and using decision-focused learning with a diffusion model to learn the distribution of scenarios.

Utsav Sadana talks at #cors2026 about how to avoid underestimating entropic risk by using parametric bootstrapping.

Gabriel Fortin-Leblanc talked at #cors2026 about a more tractable approach for solving optimization problems with decision-dependent uncertainty.

El Mehdi Er Raqabi talked at #cors2026 about how to use an optimization proxy approach to solver Benders subproblems: - Predict the dual multipliers (this is where ML replaces optimization) - Project the solution (to ensure feasibility) - Complete (a closed-form step to strengthen the cut)

Having shown that such a certificate tree exists (as a generalization of the single scenario in the case of two-stage stochastic program) but not knowing the optimization solution beforehand, Hajamini uses decision-focused learning for learning a context-dependent tree. 2/2

Mahsa Hajamini talks at #cors2026 about how to retrieve a tree of scenarios for a multistage stochastic program to obtain the true optimal first-stage decision. 1/2