Mohammed Amine Bennouna

About me

I am a 4th year doctoral candidate at the MIT Operations Research Center, advised by Prof. B. Van Parys. Prior to joining MIT, I graduated from Ecole Polytechnique in 2019 Majoring in Applied Mathematics and completed two years of preparatory program (CPGE) in Lycée Louis-le-Grand.

My research interest lies in data-driven decision-making. I strive to understand theoretically how to use most efficiently observed data to make decisions. The goal is to derive from these theoretical insights novel algorithms that tackle challenges in data-driven decision-making. My recent work focuses on designing novel robust approaches to Machine Learning problems (and more generally Stochastic Optimization) using Distributionally Robust Optimization. The goal is to understand what are the sources of overfitting (poor generalization) in data-driven stochastic optimization and design robust approaches precisely pretecting against such sources. These approaches are typically provably optimal (i.e the best that can be done with the data) when a certain out-of-sample guarantee is desired. Earlier in my PhD, I have also worked in Reinforcement Learning (RL) to introduce a novel more efficient and interepretable off-line RL framework. We have then applied this framework to COVID-19 predictions and vaccine allocation. Finally, I worked on analyzing the representation power of neural networks and contributing to explain theoretically why neural networks work so well in data-driven prediction.

---

PhD Papers

• Holistic Robust Data-Driven Decisions (paper)
  with Bart Van Parys. Working Paper.
• Learning and Decision-Making with Data: Optimal Formulations and Phase Transitions (paper , talk video)
  with Bart Van Parys. Submitted.
Winner of MIT Operations Research Best Student Paper 2022
• Learning the Minimal Representation of a Dynamic System from Transition Data (paper , talk video)
  with Dessislava Pachamanova, Georgia Perakis & Omar Skali Lami. Submitted.
• Shallow and Deep Networks are Near-Optimal Approximators of Korobov Functions (paper , talk video)
  with Moise Blanchard. ICLR 2022.
• COVID-19: Prediction, Prevalence, and the Operations of Vaccine Allocation (paper)
  with Joshua Joseph, David Nze-Ndong, Georgia Perakis, Divya Singhvi, Omar Skali Lami,
  Yannis Spantidakis, Leann Thayaparan, Asterios Tsiourvas. MSOM 2022.
  Runner-up of INFORMS Doing Good with Good OR award 2021; INFORMS ICSS Best Conference Paper Award 2021
• Near Optimal Tractable Threshold Policies for Two-stage Robust Optimization Problems
  with Omar El Housni & Vineet Goyal. Working Paper.
  Winner of Ecole Polytechnique’s 1st Prize of research internship in Applied Mathematics.

---

Undergraduate Research Projects

• Complexity of Interior-point Algorithms: Tropical Computations (see)
  with Yassine El Maazouz, Stéphane Gaubert & Xavier Allamigeon. 2018.
  This work is the continuity of the following paper: Log-Barrier Interior Point Methods Are Not Strongly Polynomial.
• On Random Minimal Factorizations of Large n-cycles (see)
  with Yassine El Maazouz & Igor Kortchemski. 2019.
• Simulating Rare Events with Hawkes Processes (see)
  with Yassine El Maazouz, Emmanuel Gobet & Martin Bompaire. 2018.

---

Teaching

• Head Teaching Assistant, MIT
  15.093/6.255 Optimization Methods, Graduate (Masters, PhDs, MBAn, MBA & others), Fall 2021. (220 students)
• Teaching Assistant, MIT
  15.093/6.255 Optimization Methods, Graduate (Masters, PhDs, MBAn, MBA & others), Fall 2020. (125 students)
• Instructor, Mathematical Olympiad
  President of the Moroccan Mathematical Olympiad society. (2016-2019)
• Instructor, Institut Bossuet
  Instructor in advanced mathematics for undergraduate students of top french classes préparatoire.

---

Talks

Holistic Robust Data-Driven Decisions
Upcoming: ORC seminar spring 2023, SIAM Conference on Optimization 2023.
Past: INFORMS Annual meeting October 2022, MoroccoAI webinar October 2022, ICCOPT July 2022, ORC Student Seminar October 2022.
Learning and Decision-Making with Data: Optimal Formulations and Phase Transitions
Past: ORC Student Seminar October 2021; INFORMS annual meeting October 2021 (see video).
Minimal Representation Learning: Toward more Interpretable and Efficient Offline Reinforcement Learning
Past: INFORMS annual meeting November 2020; ORC Student Seminar April 2021 (see video); MSOM Conference June 2021; INFORMS Healthcare Conference July 21st 2021 ; INFORMS annual meeting October 2021.
The Representation Power of Neural Networks
Past: MIT SIAM student seminar December 2020 (see video); ORC Student Seminar March 2021; Poster session ICLR 2022.