Multi-agent Learning of Causal Networks in the Internet of Things

Stefano Mariani, Pasquale Roseti, Franco Zambonelli

Abstract: In Internet of Things deployments, such as a smart home, building, or city, it is of paramount importance for software agents to be aware of the causal model of the environment in which they operate (i.e. of the causal network relating actions to their effects and observed variables to each other). Yet, the complexity and dynamics of the environment can prevent to specify such model at design time, thus requiring agents to learn its structure at run-time. Accordingly, we introduce a distributed multi-agent protocol in which a set of agents, each with partial observability, cooperate to learn a coherent and accurate personal view of the causal network. We evaluate such protocol in the context of a smart home scenario and for a two-agents case, showing that it has superior accuracy in recovering the ground truth network.