Antoine Bergerault

Master student in Data Science at the Swiss Federal Institute of Technology in Lausanne (EPFL)

Second-year Master's student in Data Science. Particularly interested in game theory, probability, numerical methods, artificial intelligence and related subjects.

Linkedin Github

Preference-based social welfare maximization 2024

Semester research project

We tackle the problem of the decentralized optimization of a global objective by combining mechanism design and preferential bayesian optimization.

Designing utilities in games so as to maximize social welfare objectives is a central problem in many applications. We study how this can be done when the social welfare is only accessible via preference feedback provided by a central authority. In particular, we present SDO-PBO, a novel algorithm that we analyze for the case of DR-submodular welfare functions, found in practice in allocation or congestion games. We pave the way on how it can be implemented for the cooling of computer clusters or the optimization of bike-sharing systems, before further discussing the limitations of our work and potential future research directions.

By combining preferential bayesian optimization with mechanism design, this project explores how a central agent can design rules of a game to maximize a social welfare function using a history of preference feedback. In this report, we develop and analyze a novel solution for this question. We first extend the work done by previous work to a game-theoretic setting, and then derive SDO-PBO, an algorithm that we prove to be efficient in the optimization of DR-submodular welfare functions.

Optimizing For Off-Road Navigation 2023

16745 Group Project

We studied ways to plan and track an optimal trajectory for off-road driving.

This document is summarizing our project on off-road car navigation from cost maps of real-world data. Based on data collected by the AirLab at Carnegie Mellon University, we were able to design a trajectory planner using RRT# and determine optimal controls to follow this trajectory. We experiment with iLQR and MPPI to track the reference trajectory using simple bicycle model dynamics.

Autonomous off-road driving on complex terrain is a challenging task. In this project, we were interested in the trajectory optimization of a vehicle in an environment derived from real-world data. To describe the vehicle we used bicyle model dynamics, and the data we used for the maps has been collected with a Yamaha vehicle at Gascola, PA by AirLab researchers. This laboratory researchers have generated all maps using various sensors such as cameras, IMU, and Lidar. Once the observations and cost maps are generated, we can derive a near-optimal trajectory by the means of a sample-based planner. In order to track the trajectory in this complex environment, we wanted to compare a sampling-based method called Model Predictive Path Integral (MPPI) to the classic iLQR-LQR pair.

Handling Heterogeneous Clients in Federated Learning 2023

18460 Group Project

Many applications deal with networked devices with heterogenous computational capabilities. In this reproducibility study, our goal is to implement and study the main algorithm of the FedHybrid paper.

Federated learning is a trending research area where we divide training between multiple devices possessing their own local datasets. It is a promising solution to overcome data security and privacy concerns. When communicating, clients involved in the training are not explicitly sharing their dataset information, but rather the result of their local updates on the trained parameters. By aggregating all the updates together, a central server would be able to leverage all these local updates and derive global consensus parameters.

Many past studies have gone over the case of homogeneous clients, where classical first order or second order updates have been proved to exhibit different convergence guarantees. Oftentimes, clients within the network do not have the same computational abilities leading to wasted resources. The FedHybrid paper came up with a solution to perform federated learning in these heterogeneous settings. In this study, we summarize how their method works, and show how we were able to implement it using Python in order to reproduce their results.

Quantum-proofing encryption systems 2019

Supervised Personal Work (Baccalaureate)

This work carried out with Ambroise Grandjean and with the supervision of some professors allowed us to give an answer to the following problem: do quantum computers challenge the current encryption systems?

Here is a transcript of the conclusion we reached:

Cryptography is currently one of the pillars of computing and secure data transfer. The most common encryption systems use mathematical theorems to base their security on technological limits. For example, the size of our encryption keys is standardized, making the time to decrypt encrypted messages far too large for current technology. Based on this observation, current encryption systems are considered secure and are used in all areas of daily life [...]. With the great advances in quantum physics and technology in particular, [quantum computers] will theoretically have the computing power to bypass these security systems. However, techniques could be used to protect the data of this new technology but currently, few are already implemented because they remain quite experimental like quantum cryptography. Hashing is used quite a lot but the problem of irreversibility does not allow it to be applied everywhere. For the moment, the quantum computer does not present any real danger for data security, but the race is on [...]. A societal problem could then arise, so we must put in place today systems that will allow us to protect ourselves from tomorrow's attacks.