Modeling the structure of collective intelligence

Abstract

The human problem solution process has attracted an increasing amount of interest among educators, managers, computer scientists and others. However, the discussion of the subject has suffered from the lack of stochastic tools to quantitatively capture both the subtler steps of the problem solution process and the diversity of human thinking. In order to stochastically model the human problem solution, this thesis presents an approach referred to as "influence modeling," that attempts to describe how an individual navigates from one random memory chunk to another related memory chunk, and how a group of people randomly remind one another of memory chunks that could be individually uncommon. As an application of influence modeling, this thesis shows how groups play "20-questions" games based on a semantic space, such as ConceptNet (a common-sense database). It also investigates how groups send signals about their behavior, which are collected by embedded devices, how group interaction processes could be automatically monitored with embedded devices, how group performance could be facilitated, and how to map group behavior and performance from the macroscopic level to the microscopic level in experiments in measuring collective intelligence. The influence modeling makes it possible to understand how a group could perform better than an individual. It also allows for the monitoring of the status of the problem solution, and makes it possible to direct group interaction in more fruitful ways.