Reducing Collective Behavioural Complexity through Heterogeneity

Abstract

In this paper, the correlation between behavioural heterogeneity and behavioural complexity within groups of cooperating agents is investigated. This investigation is accomplished using the Legion system, a type of evolutionary algorithm for evolving group behaviours in which behavioural differences among agents in the group is subject to selection pressure. Two collective task domains are studied, and two types of control architecture for the agents are used. From the experiments reported here it is concluded that increased behavioural heterogeneity within a group leads to reduced average control complexity, and also that reducing the maximum size of control architectures results in the evolution of increased behavioural heterogeneity. It is argued that this correlation helps to clarify the relationship between robustness, division of labour and variation within cooperating agent populations, and also that heterogeneity can be a useful tool for robot group design.