Abstract:By "social learning," in this article we refer to mechanisms for opinion formation and decision making over graphs and the study of how agents' decisions evolve dynamically through interactions with neighbors and the environment. The study of social learning strategies is critical for at least two reasons. On one hand, it allows for a deeper understanding of the fundamental cognitive mechanisms that enable opinion formation over networks and the propagation of information or misinformation over them. On the other hand, these same learning strategies are effective for decision making by networked agents under challenging conditions, such as highly dynamic environments, nonstationary models and data, untruthful or malicious agents, sparsely connected graphs, and constrained communication. The article presents a unifying framework that covers several cases of interest, such as single-agent Bayesian learning, multiagent non-Bayesian learning, adaptive social learning, social machine learning, partial information sharing, influence discovery, and many others. The presentation highlights important limitations of the traditional social learning strategies. One limitation is the inability to track well drifting conditions. Traditional approaches lead to stubborn agents, which resist new states of information and are slow to react to changes in the environment, like an opinion that changes over time. Another limitation of the traditional strategies is that they assume perfect knowledge of the data models, which is seldom available in practice. The article illustrates recent advances that address these issues. We show how to endow multiagent networks with adaptation abilities and how to build social machine learning solutions that learn the necessary models directly from the data. These are fundamental steps toward the construction of socially intelligent networks, capable of exploiting cooperation and diversity across the agents to guarantee reliable learning performance under nonstationary, heterogeneous, and uncertain environments.

Socially Intelligent Networks: A framework for decision making over graphs