Approche multi-agent combinant raisonnement et apprentissage pour un comportement éthique

The need to imbue Artificial Intelligence algorithms with ethical considerations is more and more present. Combining reasoning and learning, this paper proposes a hybrid method, where judging agents evaluate the ethics of learning agents’ behavior. The aim is to improve the ethics of their behavior in dynamic multi-agent environments. Several advantages ensue from this separation: possibility of co-construction between agents and humans; judging agents more accessible for non-experts humans; adoption of several points of view to judge the same agent, producing a richer feedback. Experiments on energy distribution inside a Smart Grid simulator show the learning agents’ ability to comply with judging agents’ rules, including when they evolve.