Role

Machine learning for edge networks and distributed sensors

My Master 2 research internship, hosted by Université Paris-Saclay in partnership with Green Communications. I started by using machine learning to configure the company's wireless mesh routing software, then turned a game-theoretic framework for clustered federated learning into a tested simulation whose results went into a paper I co-authored.

Role
ML Research Intern (Master 2)
Organisation
Université Paris-Saclay, with Green Communications
Period
November 2025 to July 2026
Output
Co-authored paper, submitted 2026
Stack
Python, NumPy, pandas, Matplotlib
Updated
Phase 1 · Mesh routing with Green Communications
Phase 1 · Mesh routing with Green Communications

Phase 1: configuring a mesh network with machine learning

Green Communications builds portable wireless mesh networks that work without fixed infrastructure. Their routing software has eight free settings, and every choice trades how quickly the network reacts against how much control traffic it carries. I took over an approach that tried to predict a full configuration from a single target value, showed why it could never work (many different configurations give the same result), and replaced it with lookup-table policies and a forward model (R² ≈ 0.97).

The company tested my configurations on networks of 10, 28 and 104 nodes. At 104 nodes the first batch fell short of their baselines. I traced the cause to thresholds that kept too many weak links, fixed it with domain-informed constraints and sent a corrected batch back for testing.

A model that looks convincing on a training set is not the same as one that behaves well on a real network.
Phase 2 · Clustered federated learning for soil sensors
Phase 2 · Clustered federated learning for soil sensors

Phase 2: clustered federated learning for soil-moisture sensors

The second phase, with my supervisor Dr. Kinda Khawam, became the core of the internship. A newly installed soil sensor has too little data to learn on its own, and one shared model fits sensors in different soils badly. Dr. Khawam's framework groups similar sensors, trains one federated model per group and uses game theory to admit new sensors online. My job was to turn that theory into a working simulation and test every claim.

I built a synthetic testbed of 80 soil-moisture sensors, implemented all three layers of the framework and followed the evidence when the theory needed correcting, most clearly when the admission rule, implemented exactly as written, sent nearly every new sensor to the same group.

  • Cold start: clustered learning had the lowest error at every data size, with an RMSE of 2.43 after 12 readings against 3.18 for local models and about 3.2 for a single global model.
  • Groups learned from data: the discovered groups beat grouping by soil type once sensors had 20 or more readings.
  • Fast and stable: both clustering solvers found the same five groups in well under a second, and online admission converged every time.

Every result was averaged over 20 simulations and checked in an independent audit. The work went into a paper I co-authored: K. Khawam, S. Lahoud, R. Touma and S. Martin, A three-layer game-theoretic framework for soil-moisture clustered federated learning, submitted in 2026.

What I took away

This internship taught me to work independently on open research questions, to treat a negative result as information, and to check every number before anyone else sees it. It also confirmed that I want to keep doing research, and I hope to continue with a PhD.