Data-Driven Supervision of Autonomous Systems

Modern software systems execute in increasingly dynamic settings, and their objectives are in constant motion. In order to preserve their adequacy and effectiveness within an evolving environment, software and its requirements need to adapt to change. In this dissertation, we propose a data-driven supervision framework for the automatic run-time revision of requirements, so to ensure the achievement of system-level objectives in dynamic settings. We focus on the supervision of multi-agent systems (MASs), collections of interacting autonomous agents, such as autonomous cars on smart roads. In multi-agent systems, agents’ internals are typically unknown to the other agents and to the MAS designer. Norms are often employed as a means for controlling and coordinating the agents' behavior without over-constraining their autonomy. We use norms to characterize requirements for the behavior of the agents in the system, and we use sanctions as a deterrence mechanism to discourage agents from violations. The proposed supervision framework employs a general architecture for system self-adaptation, described as a closed control-loop. At run-time, the system is monitored and execution data is collected in different operating contexts. The collected data is used to learn statistical correlations between the achievement of the system's objectives and the satisfaction of the requirements in the different operating contexts. The learnt information is applied to automatically assess the validity of the assumptions made at design-time, and to automatically synthesise new requirements and sanctions when there is evidence that the current ones are not effective