In this paper, a genetic algorithm (GA) is used to discover interaction rules for a cellular self-organizing (CSO) system. The CSO system is a group of autonomous, independent agents that perform tasks through self-organization without any central controller. The agents have a local neighborhood of sensing and react only to other agents within this neighborhood. Their interaction rules are a simple set of direction vectors based on a flocking model. The five local interaction rules are assigned relative weights, and the agents self-organize to display some emergent behavior at the system level. The engineering challenge is to identify which sets of local rules will cause certain desired global behaviors. The global required behaviors of the system, such as flocking or exploration, are translated into a fitness function that can be evaluated at the end of a multi-agent based simulation run. The GA works by tuning the relative weights of the local interaction rules so that the desired global behavior emerges, judged by the fitness function. The GA approach is shown to be successful in tuning the weights of these interaction rules on simulated CSO systems, and, in some cases, the GA actually evolved qualitatively different local interaction “strategies” that displayed equivalent emergent capabilities.
