Multi-agent Online Planning Architecture for Real-time Compliance

Due to the recent rapid introduction of AI technologies into society, we face new risks related to AI. Therefore, it is very important to let AI be compliant with legal and ethical norms to reduce such risks. In this paper, we propose a method of compliance mechanism of AI agent planning in a multi-agent setting. We formalize legal norms as hard constraints which must be satisfied and ethical norms as soft constraints which should be satisfied as much as possible. Moreover, our mechanism can handle real-time exogenous environmental change by a replanning mechanism. We believe that our architecture is an important step towards "trustworthy AI".