A multiagent system (MAS) comprises a set of autonomous agents that interact in a shared environment [9] . For example, a smart traffic system is a MAS that includes autonomous agents like cars, traffic lights, etc. The objectives of such a system include ensuring that each agent reaches its destination timely, and that the number of accidents is minimized. For a MAS to achieve its system-level objectives, the complexity and unpredictability of the agent interactions and of the environment must

more »

... be taken into account. When engineering such systems, the available knowledge of these dynamics is only partial and incomplete. Therefore, MASs need to be regulated at runtime. Norm enforcement is a prominent mechanism for controlling and coordinating the runtime behavior of the agents in a MAS without over-constraining their autonomy [1, 8] . Norm enforcement via sanctions is traditionally contrasted with norm regimentation, which prevents the agents from reaching certain states of affairs. For example, in a smart traffic system, a regimentation strategy could be to close a road to prevent cars from entering the road, while a sanctioning strategy could be to impose sanctions on cars that do enter the road. Existing research has studied the offline revision of the enforced norms, and proposed logics that support norm change [2, 7], and studied the legal effects of norm change [6] . In previous work [4] we have proposed a framework for engineering normative MASs that, using runtime data from MAS execution, revises the norms in the MAS to maximize the achievement of the system objectives. In such work, we made the simplistic assumption that norms are regimented and we introduced algorithms for switching among alternative norms. In this paper, we make a step forward toward the engineering of normative multiagent systems and we propose a regulatory mechanism that relies on norms with sanction. In our approach, we automatically revise the sanctions that are employed to enforce the norms. To do so, we first interpret-through a Bayesian Network-runtime execution data in terms of how well certain norms contribute to the achievement of the system-level objectives in different operating contexts. Then, we suggest a revision of the sanctions using two different heuristic strategies, called Synergy and Sensitivity. The two strategies leverage the knowledge learned from runtime execution data and the knowledge about the