Adrien Bolling

Doctoral researcher in AI · SnT, University of Luxembourg

Adrien Bolling

I study reinforcement learning for decisions that have more than one objective and more than one agent. My main application is industrial maintenance: which technician repairs which machine, and what the team learns from it.

A good dispatch policy keeps the line running today. It must also let the technicians grow their skills, and it must not exhaust them. I build simulators, knowledge models, and learned policies for this problem in the SerVal group.

A technician's knowledge grid, as in ONGOING. Each repair adds experience around its task, and related tasks gain a part of it. Unused skills fade slowly.

Research

Four lines of work. They share one question: how do you learn a policy when people are part of the system?

01

Human-centric maintenance

Technician allocation is a human resource allocation problem. Most methods treat skills as fixed. In my work, skills grow with each repair, fatigue builds up, and experts retire. The FactoReal simulator models these dynamics.

Bolling et al., preprint 2026

02

Multi-objective, multi-agent RL

A dispatch policy must balance production, knowledge growth, and workload. I train set-based Transformer policies with PPO for this setting. One policy accepts any number of technicians and machines.

Bolling & Kubler, MODeM 2024

03

Knowledge modelling

ONGOING turns free-text maintenance tickets into human-readable knowledge grids. Experts can audit them, and models can use them as features. On 55 real tickets, expert dispatchers rated its recommendations above scalar and vector models.

Bolling & Kubler, PHM Asia-Pacific 2025

04

Time-series augmentation

Industrial data is small and unbalanced. ASCENSION generates synthetic time series in the latent space of a VAE and expands each class in a controlled way. It improves classifiers on 102 UCR datasets.

Joubaud et al., TMLR 2026

Interactive · about 5 minutes

Dispatch: can you keep the plant running after your best technician retires?

You run a maintenance crew for ten days. Machines fail, and you send technicians to repair them. Marta, your veteran, retires on day 5. Your score is compared with three dispatch policies on the same failures.

Start a shift

Publications

Also on Google Scholar. My name is in bold.

  1. 2026Journal

    ASCENSION: Autoencoder-Based Latent Space Class Expansion for Time Series Data Augmentation

    Dorian Joubaud, Matthieu Olekhnovitch, Adrien Bolling, Evgeny Zotov, Sylvain Kubler, Maxime Cordy, Mike Papadakis, Yves Le Traon

    Transactions on Machine Learning Research (TMLR), 2026

    OpenReview
  2. 2026Preprint

    Knowledge-Aware Technician Allocation: The Long-Term Impact of Technician Upskilling

    Adrien Bolling, Sylvain Kubler, Marcelo Luis Ruiz-Rodríguez, Yves Le Traon

    Preprint, SSRN, 2026

    FactoReal simulates a maintenance crew whose knowledge and fatigue change with each assignment. A hybrid-token Transformer policy (HTT-RL) is compared with thirteen baselines in five scenarios. The results show that human-centric objectives can improve both the sustainability of the workforce and the long-term production.

    Preprint Code (Zenodo) Play the game
  3. 2025Conference

    ONGOING: A Human-readable, Model-enriching, Continuous Technician Knowledge Modeling Framework

    Adrien Bolling, Sylvain Kubler

    Asia Pacific Conference of the PHM Society, vol. 5, no. 1, 2025

    ONGOING turns maintenance tickets into knowledge grids that track the experience of each technician. Two mechanisms model the transfer of knowledge between technicians and its diffusion to related tasks. It runs on 50,000 real tickets and 100 technicians.

    Paper DOI Code
  4. 2024Workshop

    Knowledge-aware and Learning-focused Multi-Objective Multi-agent Reinforcement Learning for Maintenance Technician Assignment

    Adrien Bolling, Sylvain Kubler

    Multi-Objective Decision Making Workshop (MODeM 2024) at ECAI 2024, Santiago de Compostela

    PDF

Software

Open-source code from my research and for my daily work.

Timeline

  1. ASCENSION is accepted at TMLR.
  2. The code of the technician upskilling study is on Zenodo. The preprint is on SSRN.
  3. ONGOING appears in the PHM Asia-Pacific proceedings.
  4. Our paper on knowledge-aware MOMARL appears at the MODeM workshop at ECAI 2024.
  5. I start my PhD at SnT, University of Luxembourg.