Distributed Coevolutionary Design of Velocity Controllers for Swarm Intelligence Robots in Flocking and Entrapping Tasks

Abstract

This study combined evolutionary algorithm and reinforcement learning to propose a new automated design method for generating swarm robots velocity controller model. It alternately evolves heterogeneous swarm and homogeneous swarm through a gene expression programming method that introduces reinforcement learning, and assembles function nodes and leaf nodes into new mathematical formulas during the evolution process. The method enable to realize the effect of swarm robots emerging to perform swarm tasks such as flocking and entrapping. What’s more, a new swarm rule was discovered during the evolution process, which is used to realize the flocking of swarm robots at any angle. The experimental results show that the swarm motion controller automatically generated by the model has high task completion efficiency and strong generalization.

Keywords: swarm intelligence, robots, evolutionary computation, multi-agent systems, gene expression programming, reinforcement learning