Optimizing group learning: An evolutionary computing approach

摘要

We study groups of interacting agents who are trying to discover probabilistic truths on the basis of sequentially provided evidence and information about the belief states of other group members. The main research question is which combination of epistemic principles—combinations of an evidential update rule, a rule for determining peerhood, and a rule for aggregating probability functions—such groups should adopt to strike the best balance between being fast and being accurate, where the former is understood by reference to how long it takes before a majority of the group assigns a high probability to the true hypothesis, and the latter by reference to the average Brier penalty incurred by the group. The main methodology to be used is that of agent-based optimization, which is a specific form of evolutionary computing. We implement this methodology in a generalization of the Hegselmann–Krause model. In the end, we are able to identify optimal procedures for taking into account both direct evidence and information about one's peers' beliefs. At the same, we note that optimality for such procedures is dependent on context.