Learning Risk Preferences Through Social Interaction: An Active Inference Approach

Social learning involves the ability to make assumptions about a person's attitudes and preferences on the basis of their observed behaviors. These characteristics are then used to make predictions and influence how one interacts with individuals in various social contexts. In this research, we used an active inference framework to explore how humans can assess another person's risk preferences. The active inference learner uses a probabilistic generative model to model others' decision making process and then employs a Variational Bayesian method based on the free-energy principle to invert the generative model. On a trial-by-trial basis, the observer updates the posterior estimate of the other's hidden characteristic via social prediction error, a mismatch between what we predicted and what they chose. The model begins with a high learning rate and gradually reduces it as more trials occur. This reflects that humans learn from examples sequentially, exploring and using the information as they progress. As more data are accumulated from someone who follows a consistent decision-making process, the observer's faith in his own judgments should grow, and he should be less affected by each new observation. Understanding others' risk preference is a complex example of social reasoning. We employed a sequential scenario where an observer watched the other person's choices between a high-risk gamble and a guaranteed small reward. We discovered that the inferential uncertainty that the learner has in trying to estimate the agent's risk attitude, is largely based on the prediction uncertainty accumulated by the observer when attempting to guess the agent's choices.