Research (Publications)

Reinforcement learning (RL) is becoming more prevalent in practical domains with human implications, raising ethical questions. Specifically, multi-objective RL has been argued to be an ideal framework for modeling real-world problems and developing human-aligned artificial intelligence. However, the ethical dimension remains underexplored in the field and no survey covers this aspect. Hence, we propose a review of multi-objective RL from an ethical perspective, highlighting existing works, gaps in the literature, important considerations, and potential areas for future research.

Reinforcement learning (RL) is becoming more prevalent in practical domains with human implications, raising ethical questions. Specifically, multi-objective RL has been argued to be an ideal framework for modeling real-world problems and developing human-aligned artificial intelligence. However, the ethical dimension remains underexplored in the field, and no survey covers this aspect. Hence, we propose a review of multi-objective RL from an ethical perspective, highlighting existing works, gaps in the literature, important considerations, and potential areas for future research.

This paper presents a novel Multi-Objective Reinforcement Learning approach to settle dilemmas, by putting humans in the loop.

Paper published in the Journal of Open-Source Software, alongside the source code for our ethical-smart-grid simulator. This simulator focuses on ethical behaviours within a Smart Grid, and is based on Gym (Reinforcement Learning).

Journal paper published at the French Artificial Intelligence Open Journal (Revue Ouverte d’Intelligence Artificielle). This works extends previous works, especially the conference paper published at JFSMA 2021.

Paper presented at the Autonomous Agents and Multiagent Systems conference. It presents the AJAR framework, which uses argumentation-based judging agents to provide rewards for Reinforcement Learning agents, according to one or several moral values. This “judgment of ethics” is used to nudge the learning agents towards an “ethical behavior”, that is, a behavior aligned with the given moral values.

Preprint describing two Reinforcement Learning algorithms (Q-SOM and Q-DSOM) I have developped. They focus on continuous and multi-dimensional observations and actions, and adaptation to changes in the environment.

This paper presents how Artificial Intelligence could be used to help humans in their ethical decision-making tasks.

PhD thesis on Learning behaviours aligned with moral values in a multi-agent system: guiding reinforcement learning with symbolic judgments, realized at the LIRIS lab, under the supervision of Professor Salima Hassas (LIRIS), Professor Olivier Boissier (LaHC), and Dr. Mathieu Guillermin (UCLy).

A technical report that served as a preparatory document for a tutorial that we presented at PFIA (French Platform on Artificial Intelligence) 2020, organized by the AFIA (French Association for Artificial Intelligence).

Paper on the ability to use a symbolic reasoning approach to judge neural learning agents, in order to reward them appropriately with respect to their ’ethical’ behavior, combining both approaches in a Hybrid method.

Paper presented at the AI, Ethics, and Society conference. It presents a novel, hybrid, method to learn “ethical behaviors” by combining symbolic judgments with a Reinforcement Learning algorithm.

Paper presented at the Explainable Agency in AI Workshop, hosted by the 35th AAAI conference. It presents a quick review of explanation in the social sciences and in the AI communities, and proposes to move towards a User eXplainable Artificial Intelligence model (UXAI) to place the user at the center of the explanation process.

Paper on adaptive learning of ethical behaviors I presented at JFSMA (French Days of Multi-Agent Systems).

Poster about my Master thesis, at an AI Autumn school.

Master’s thesis on Evolutive learning of ethical behaviors, realized at the LIRIS lab, under the supervision of Professor Salima Hassas (LIRIS) and Professor Olivier Boissier (LaHC).