Post-doc in “ethical Multiagent Reinforcement Learning: a Normative-base Approach”

Joining Mines Saint-Étienne means committing to an institution where **science and innovation build a more sustainable future**. It is a school of excellence where everyone has the **opportunity to unlock their full potential** and **contribute to tackling the challenges of tomorrow**. Ranked among the top engineering schools in France and recognized worldwide, our school, a member of the _Institut Mines-Télécom_, educates the talents of tomorrow while actively addressing major industrial, digital, and environmental challenges. By joining us, you become part of a community of 500 members of staff, 2,500 students and take part in an ambitious project: combining academic excellence, cutting-edge research, and positive societal impact. The _Institut Mines-Télécom_ brings together France’s leading _Grandes Ecoles_ to tackle major industrial, digital, energy, and environmental challenges. With its eight public _Grandes Écoles_ and two affiliated Graduate Schools, it is the leading public institute dedicated to engineers and managers. Together, we imagine and build a sustainable future by educating the leaders who will shape tomorrow’s transitions. **What you expect from you** As a **Post-doc**in **ETHICAL MULTIAGENT REINFORCEMENT LEARNING**, you will be at the heart of our missions in research, and innovation, assigned to the **Laboratory of Informatics, Modelling and Optimization of the Systems (LIMOS) UMR 6158 CNRS** and hosted at the **Institute Henri Fayol Education and Research Centre**. Within this centre, you will combine your passion for science and for societal impact. The **Institute Henri Fayol** focuses on current transformations in the digital, ecological and industrial transitions that are at the heart of the efficiency, resilience and sustainability of industry and territories. It develops a multi-disciplinary strategy combining strong skills in mathematical and industrial engineering, computer science and intelligent systems, environmental and organizational engineering, and responsible management and innovation, in conjunction with the EVS UMR 5600, LIMOS UMR 6158 and COACTIS research units. The importance Artificial Intelligence (AI) systems are gaining in our daily lives makes it a crucial and pressing matter to ensure that they are in line with (moral) values and respect social and legal norms. These systems, by interacting with humans and, more generally, being immersed in our societies, have a direct impact on our lives. This urges AI researchers to develop more ethically-capable systems, shifting from ethics in design [2] to ethics by design with "explicit ethical agents" [4] able to behave ethically thanks to the integration of reasoning on and learning of ethics. Reinforcement learning (RL) methods enable agents to learn making decisions, but they do not guarantee safety or compliance with ethical values or legal and social norms. Safe reinforcement learning (SRL) [3] has been proposed to ensure reasonable system performance and respect to safety constraints during the learning or deployment processes. The Shield RL [1] monitors the environment and prevents the execution of actions that would violate formally defined safety constraints or norms. The Norm Guided RL [5] uses a normative supervisor that evaluates the potential agent's actions against a normative system and informs the decision-making process of the agent, which may still choose to perform an action that violates safety constraints or norms. In the context of the ACCELER-AI project, we extended the Norm Guided RL proposed in [5] to enable expressing norms declaratively and in ordering preference using Answer Set Programming (ASP). By being declarative, those norms may be easily changed either by human beings or by dedicated agents. Our preliminary results in a simple Pac-Man scenario replicates previous works and corroborates their conclusions that influencing directly the learning process based on the information about the violation of ethical constrains leads to better outcomes than only constraining agents' action space at run-time. The post-doc will investigate further the Norm Guided RL approach in a different scenario and contrast the advantages and disadvantages of this approach with the Reward Shaping approach in RL [6]. Once concluded, the post-doc will extend the proposed approach to bounding ethical behaviours in the context of Multi-Agent Reinforcement Learning (MARL) and demonstrate its effectiveness in a prototypical demonstrator in the domain of mobility or smart grid. Reporting to the ACCELER-AI project coordinators and working in collaboration with the other project partners, post-doc main tasks will be: - Design and perform experiments of the Norm Guided RL approach in a single agent context considering different scenarios (e.g., harvesting [7]) contrasting it with Reward Shaping. - Elaborate a state of the art on mechanisms for bounding agents learning in the context of MARL. - Design novel mech