Abstract

In social learning the way in which agents represent their beliefs motivates and constrains both how they learn individually from the environment and socially from one another. Assuming that agents can only hold beliefs drawn from a finite set of possible belief states, in this paper we investigate the effect that varying the number of those belief states has on the efficacy of social learning. To this end we propose an evidence tokens model for social learning, in which agents transfer tokens between competing hypotheses on the basis both of evidence that they receive directly and of information received from their peers. Using agent-based simulations and difference equations we show that this model is effective in social learning for boundedly rational agents and scales well to the case where there are multiple hypotheses under consideration. We show that varying the number of belief states (as determined by the number of evidence tokens available) has a clear effect both on accuracy and on the time taken for the agent population to reach agreement about which hypothesis is true, so that the optimal belief granularity in social learning is strongly influenced by macro properties of the whole population governing the way that agents interact with each other and the environment.