A Biased Inferential Naivety Model for Agents’ Learning in Social Networks

We proposed a model of learning and belief formation in which a group of agents tries to learn the true underlying state of the world and make the best possible decisions. Agents with limited computational ability, in addition to receiving noisy private signals, observe the decisions of their neighbors. It is well known that Bayesian inference is very complex in social observations, especially when agents are unaware of the structure of the social network. In our model, the role of knowledge derived from the social observations of each agent is separated from that’s of her private observations in the formation of her belief. Thus, to reduce the complexity of Bayesian inference, the processing of social observations is approximated using the inferential naivety assumption. With this assumption, agents naively believe that each neighbor's decisions are based solely on his or her private observations and that their social interactions are ignored. Another important initiative in the proposed model is to eliminate herd behavior by introducing an exponential bias and reducing the weight of early social observations compared to recent observations. A number of Monte Carlo simulation experiments confirm the features of the proposed model. This includes asymptotic learning of all agents and increased learning efficiency in social networks.