Learning driven by surprises: Entropic Dynamics of Distrust and Opinions of Social Agents Systems

We study a society of agents that interact through the exchange of opinions on a set of issues. The learning dynamics between two agents is studied using entropic inference in a general framework.

For a particular agent architecture it leads to an optimal generalization algorithm in a given informational context. 

This two body interaction is used  to simulate groups. The agents learn form the opinions of other agents about issues and also learn about the level of distrust that should be attributed to other agents.

Learning is driven by the surprise brought by unexpected opinions. The blame of the surprise can be attributed either to a small stability of the opinion or to the fact that the agent sending the information has not been clearly identified as friend or foe. Relevant changes in the agent parameters occur  in the sector identified as the cause of the error in prediction.

Applications to small groups are presented. A group of three agents is used to model the interactions and decisions in three judge panels of the US Appellate Courts judicial system.

Agreement with empirical data (Sunstein {\it et al.}) can only be obtained if the judges are, independently of political party, trustful of each other and sure about this trust.

We find evidence that judges vote ideologically and that the interaction with judges from a different party decreases the ideological difference and with judges of the same party increase the polarization. Application to larger groups lead to spinglass like behaviour with frustration and the breakdown of balance in social relations.

speaker:
Nestor Caticha
Instituto de Física
Universidade de  São Paulo