Learning Algorithms for Designing Undeceivable Policies to Foster Sustainable Behavior

Personalized Demand-Side Mitigation Strategies (PDSMS) encourage agents to make sustainable choices.[IE21] A regulator learns agents' preferences by observing their choices, and adapts signals, e.g., incentives or prices.[Ar18] State of art PDSMS[Ar23,As21] are based on Random Utility Theory (RUT), assuming agents are honest, making choices to maximize their utility.[Be19,§3.1] We study instead the case where agents may be deceptive, making choices to manipulate the regulator and get favorable signals. Our objective is to answer the following research questions: Under which conditions deceptive agents cancel-out the benefits of PDSMS? How to make PDSMS robust to them? This remains an open question, highlighting the novelty of our project. We build our approach on recent advances in AI & Game Theory,[Gan20,Xu21] but our originality is that we will explicitly model the regulator's learning process (missing so far) and show that it can deter agents from deceiving. To this aim, we will devise a novel reinforcement learning formulation rooted in RUT and resort to Mechanism Design to make PDSMS robust to deception. [Bo15] Lab experiments will validate our findings. PDSMS can contribute to sustainability (e.g., reducing pollution up to 40% [IP23,IP22]). The theoretical framework resulting from our project can unlock their full potential. École doctorale : Ecole Doctorale de l'Institut Polytechnique de Paris Direction : Andrea ARALDO Financement : Allocation doctorale AMX