A developmental approach to evolving scalable hierarchies for multi-agent swarms

  • Authors:
  • Terence Soule;Robert B. Heckendorn

  • Affiliations:
  • University of Idaho, Moscow, ID, USA;University of Idaho, Moscow, ID, USA

  • Venue:
  • Proceedings of the 12th annual conference companion on Genetic and evolutionary computation
  • Year:
  • 2010

Quantified Score

Hi-index 0.00

Visualization

Abstract

In this paper we present requirements for a successful learning approach for large scale multi-agent swarms: individual performance, cooperation and robustness. We then present a developmental, evolutionary approach for evolving hierarchical control structures for large (100-1000 agent), multi-agent swarms that addresses these requirements. Although hierarchical, the control structure does not suffer from single point of failures as do many hierarchical structures. The approach is tested on a novel problem for which a fully distributed swarm performs poorly. The results show that for some problems using an evolved control hierarchy to guide the agents leads to significantly better performance and scaling properties than fully distributed swarms using standard, simple behavioral rules. This research suggests that hierarchies are an important organizing feature to be considered for large multiagent tasks and that the efficient automated discovery of a hierarchy is an important research objective.