coconut banner
coconut banner coconut banner coconut banner

Welcome to Coconut

Coconut is a research group in computer science at the LIRMM laboratory in Montpellier, France. We are interested in artificial intelligence, with a primary focus on constraint programming.

Research topics

Foundations of constraint programming

- Propagation, search strategies, spatio-temporal constraints, interval constraint programming, complexity

Constraints and learning

- Constraint acquisition, learning for search

Applications

- Digital agriculture, robotics, automotive industry

Members

Permanent Members

Students

  • Areski HIMEUR, PhD Student
  • Verlein RADWAN, PhD Student
  • Yoan THOMAS, PhD Student

Other (Postdocs...)

  • Zouhayra AYADI, ATER
  • Charles VERNEREY, Postdoc
You can also see our former members here:

Talks and Seminars

Talks and Seminars
Élicitation des préférences pour des problèmes multi-objectifs
2025-05-16 15:30
E323
Marianne Defresne
De nombreux problèmes d'optimisation combinatoire réels ont plusieurs objectifs, souvent en conflit, comme la qualité d'un produit, son prix et son empreinte carbone. Une manière efficace d'optimiser ces objectifs simultanément est de les agréger au sein d'une seule fonction objectif via une somme pondérée. La difficulté est de définir les poids, représentant l'importance de chaque objectif. Une approche est l'élicitation de préférence, où on présente à un décideur humain des paires de solutions à comparer. Après un aperçu des méthodes existantes, je présenterai une approche pensée pour interagir avec un humain : les interactions se font en temps réel, et le décideur peut être inconsistant dans ses réponses. J'aborderai enfin quelques cas concrets d'application, comme la conservation de la biodiversité.
Solving Games on Networks
2025-05-07 15:30
E324
Amnon Meisels
Multiagent decision-making settings feature a tension between the individual interests of agents and the promotion of a common good. These typical situations are often modeled as games. Games played among a large number of agents, on a social network, where the utility of an agent depends on the joint action of its neighbors – are games on networks (GoNs). In a GoN, all agents are assumed to be selfish. Typical optimization approaches in MAS have agents collaboratively execute distributed incomplete search algorithms. Due to the game-like strategic character of agents in a GoN, such a fully cooperative technique seems inappropriate. We propose to take a different stand, a game-theoretic one, in multi-agent games on networks. All agents are assumed to be selfish, and their coordination throughout the execution of the proposed incentive-based distributed algorithm is achieved via calls to a truthful mechanism. The resulting method guarantees the convergence of the system towards a stable state – even for games that do not have an equilibrium state in their original form. This final state is guaranteed to be at least as efficient as the initial state of the system. An experimental evaluation on various types of GoNs demonstrates the performance of the algorithm and shows several interesting results. The method outperforms a former incentive-based algorithm on Public Goods Games on networks. Efficiency and stability are guaranteed due to payments to the truthful mechanism and these payments are negligible compared to the overall gain of the system.
Learning logic programs with constraint programming
2024-11-22 15:30
Salle Séminaire
Céline Hocquette
Inductive logic programming (ILP) is a form of program synthesis where the goal is to learn logic programs that generalise training examples. The main challenge in ILP is to efficiently search through large hypothesis spaces (the set of all possible programs). In this presentation, I will describe a constraint-based approach to ILP that discovers constraints to prune the search space. I will then describe a constraint-based decomposition approach to improve learning performance.
Using Large Language Models to Improve Query-based Constraint Acquisition
2024-11-15 15:30
E323
Christian Bessiere
Most active constraint acquisition systems suffer from two weaknesses. They require the explicit generation of the set of potential constraints (the bias), whose size can be prohibitive for practical use of these systems, and the answers to queries contain little information. In this paper, we introduce AcqNogoods, an active learning schema that does not require the construction of a bias. We then propose LlmAcq, an active learning system that incorporates a Large Language Model component in the AcqNogoods schema. LlmAcq interprets the user's answers given in natural language, leading to more informative communication. As our experiments show, the non requirement of a bias in extension combined to the higher-level communication with the user allow LlmAcq to learn constraints of any arity and to dramatically decrease the number of queries.

Projects

AudioMobility 20302024-2026

Le projet AM 2030 permettra aux constructeurs automobiles de disposer de leur propre application audio embarquée, quel que soit le système d’exploitation. Ils pourront déployer une expérience globale audio et offrir les meilleurs contenus et services proactifs aux automobilistes. Il se positionne comme un véritable compagnon de route qui aidera le consommateur à adopter des comportements éco-responsables : auto-diagnostique du véhicule et rapport de maintenance, conseil quant à la conduite et l’usage des équipements à bord. Partenaires : ETX Studio, CONTINENTAL Automotive, Université Fédérale Toulouse, École Polytechnique

Contact: Christian Bessiere

TAILOR (Developing the scientific foundations for Trustworthy AI through the integration of Learning, Optimisation and Reasoning)2020-2024

The purpose of the EU Project TAILOR is to build the capacity to provide the scientific foundations for Trustworthy AI in Europe by developing a network of research excellence centres leveraging and combining learning, optimisation, and reasoning. These systems are meant to provide descriptive, predictive, and prescriptive systems integrating data-driven and knowledge-based approaches. Artificial Intelligence (AI) has grown in the last ten years at an unprecedented pace. It has been applied to many industrial and service sectors, becoming ubiquitous in our everyday life. More and more often, AI systems are used to suggest decisions to human experts, to propose actions, and to provide predictions. Because these systems might influence our life and have a significant impact on the way we decide, they need to be trustworthy. How can a radiologist trust an AI system analysing medical images? How can a financial broker trust an AI system providing stock price predictions? How can a passenger trust a self-driving car? These are fundamental questions that require deep analysis and fundamental research activity as well as a new generation of AI talents who are skilled in the scientific foundations of Trustworthy AI, who know how to assess, and how to design, trustworthy AI systems. Some of the current issues related to lack of trust in AI systems are a direct consequence of the massive use of black-box methods relying only on data. We need to define the foundations of a new generation of AI systems not only relying on data-driven approaches, but also on the whole set of AI techniques, including symbolic AI methods, optimization, reasoning, and planning.

Contact: Christian Bessiere

Hybrid AI Systems Grounded on Spatio-Temporal Reasoning2023-2028

In the proposed research project, I, together with collaborators from the Coconut team and the rest of my research network, will work on the interplay between knowledge representation and machine learning, especially in the context of attaining spatiotemporally-informed AI frameworks. Of particular importance for this integration of learning and reasoning will be the area of Qualitative Spatial and Temporal Reasoning, or QSTR for short, which is a major field of study in AI that deals with the fundamental cognitive concepts of space and time in a symbolic, human-like manner. Simply put, QSTR abstracts from numerical quantities of space and time by using qualitative descriptions instead (e.g., precedes, contains, is left of ), with applications in a plethora of areas and domains such as smart environments, intelligent vehicles, and unmanned aircraft systems. The project is tied to my Junior Professorship Chair (CPJ) and is funded by ANR and the the I-SITE program of excellence of the University of Montpellier.

Contact: Michael Sioutis

Artificial Intelligence for Health and Environment – AXIAUM2021-2024

The AXIAUM project focuses on raising innovative interdisciplinary research topics on health and environment related to artificial intelligence (AI) with the scope of developing both new methodologies and finalized research. The ambition of the programme proposed by the consortium University of Montpellier/IMT Mines Alès is to strongly link new research on AI to the existing research and the AI strategy of the consortium, reinforced by the current structuration of data science within the ISDM Montpellier Institute of Data Science.

Contact: Clément Carbonnel