Learning-based Synthesis of Social Laws in STRIPS

In a multi-agent environment, each agent must take into account not only the actions it must perform to achieve its goals, but also the behavior of other agents in the system, which usually requires some sort of coordination between the agents. One way to avoid the complexity of centralized planning and online negotiation between agents is to design an artificial social system. This system enacts a social law that restricts the behavior of the agents. A robust social law enables the agents to reach their goals while keeping them from interfering with each other. However, the problem of efficient synthesis of such laws is computationally hard, and previously proposed search techniques do not scale well. In this paper, we propose the use of graph neural networks to predict social laws from a graph-based representation of multi-agent systems. However, as this prediction can be wrong, we use heuristic search to correct possible mistakes in the network’s prediction ensuring that the produced social law is indeed robust. Our empirical evaluation shows that this approach beat the previous state-of-the-art in social law synthesis and that it can learn from an imperfect expert, even in the presence of noise.