Group Abstraction for Large-Scale Agent-Based Social Diffusion Models

In this paper an approach is proposed to handle complex dynamics of large-scale multi-agents systems modelling social diffusion processes. Based on local properties of the individual agents and their connections, groups and dynamic properties of these groups are identified. To determine such dynamic group properties two abstraction methods are proposed: determining a group invariant and approximation of group processes by weighted averaging of interactions. This enables simulation of the multi-agent system at a more abstract level by considering groups as single entities substituting a large number of interacting agents. In this way the scalability of large-scale simulation can be improved significantly. Computational properties of the developed approach are addressed in the paper. The approach is illustrated for a collective decision making model.