DESCRIPTION :
Supervision: The Postdoctoral researcher will be advised by Odalric-Ambrym Maillard from Inria team-project
Scool and Cyrille Midingoyi from CIRAD/PERSYST AIDA Unit.
Place: This position will be primarily held at the research center Inria Lille - Nord Europe, Villeneuve d'Ascq,
in the Inria team-project Scool (Sequential, Continual and Online Learning), with strong regular interactions with
CIRAD AIDA unit in Montpellier.
Keywords: Multi-armed bandits, Sequential statistics, Societal challenge., [3] Pierre Martre, Donatelli Marcello, Christophe Pradal, Andreas Enders, Cyrille Ahmed Midingoyi, Ioannis Athanasiadis, Davide Fumagalli, Dean P. Holzworth, Gerrit Hoogenboom, Cheryl Porter, H´el`ene Raynal, Andrea Emilio Rizzoli, and P Thorburn. The agricultural model exchange initiative. In IICA, editor, 7th AgMIP Global Workshop, San Jos´e, Costa Rica, 2018.
[4] William Solow, Sandhya Saisubramanian, and Alan Fern. Wofostgym: A crop simulator for learning annual and perennial crop management strategies. arXiv preprint arXiv:2502.19308, 2025
Mission confiée
Objectives
The goal of the postdoc project is to develop a robust and flexible interface between crop models and reinforcement learning (RL) to enable decision-making algorithms to interact with crop simulations. This requires bridging a significant conceptual and technical gap between monolithic crop models, in which user-defined management actions are treated as predefined input variables, and RL, which relies on state-action-reward loops to enable adaptive decision-making under uncertainty.
The core challenge, therefore, is to design a modular and generalizable methodology that embeds STICS within a standard RL framework. This will pave the way for adaptive, data-driven decision-making in agronomy and open new opportunities to optimize crop management strategies in complex, uncertain environments.
Methodology
The methodological pathway follows a generic-to-specific progression: defining a model-agnostic formalism to couple black-box PBM with reinforcement learning (RL), then implementing this abstraction in a reusable software interface, and finally instantiate and evaluate it on STICS or other PBM.
Principales activités
* O1- Define a model-agnostic formalism for RL-PBM coupling. The first step is to develop an abstract
representation of how an RL agent can interact with a PBM. It will focus on the formalization of how to
access state variables or other indicators from crop model run, how to change recommendations or manage-
ment decisions (actions) through controllable input levers (e.g. sowing date, fertilization, irrigation, cultivar
choice), and how to define reward signals that encode agronomic and environmental objectives (yield, risk,
resource use, emissions, etc.).This formalism will be expressed as a generic state-action-reward-transition
schema, compatible with standard RL frameworks, but enriched with agronomic structure (time step, crop
management, etc.). This model-agnostic abstraction will serve as a conceptual template for any mechanistic
crop model.
* O2 - Implement a generic RL-crop interface library. Building on the abstract formalism defined in O1, the
second objective is to implement a generic software interface that operationalizes this coupling, in the form
of an OpenAI Gymnasium-compatible environment layer (a "Gym-Agro" or "Gym-PBM abstraction). This
layer will expose a standardized API (reset, step, observe, reward, done) to RL agents, manage simulation
calls and time-stepping, allow the mapping between model-specific I/O and the abstract state-action-reward
schema to be specified declaratively,etc. The result will be a model-independent environment layer into
which different crop models can be plugged without changing the core RL code, simply by providing a
suitable adapter specification.
* O3 - Instantiate and evaluate the formalism on STICS The third objective is to specialize and validate the
generic methodology on STICS. Using the formalism and the "Gym-Agro" interface from O1-O2, we will
develop a STICS-specific adapter that maps STICS input variables and management options to actions,
extracts relevant biophysical and management indicators as states, defines appropriate reward functions in
line with agronomic objectives. We will demonstrate the flexibility and scalability of the approach for
diverse cropping systems and objectives.
Code d'emploi : Chargé de Recherches (h/f)
Domaine professionnel actuel : Scientifiques
Temps partiel / Temps plein : Plein temps
Type de contrat : Contrat à durée indéterminée (CDI)
Compétences : APIs, Apprentissage par Renforcement, Gerrit, Anglais, Français, Prise de Décision, Sens de la Communication, Agriculture, Algorithmes, Gestion Stratégique, Étalonnage, Recadrage, Conception d'Expériences, Scalabilité, Mathématiques, Modélisation Mathématique, Recherche Post-Doctorale, Analyse de Risques, Simulations, Etudes et Statistiques, Irrigation (Agriculture), Capacités de Démonstration, Agronomie
Courriel :
Odalric.Maillard@inria.fr
Téléphone :
0139635511
Type d'annonceur : Employeur direct
