Neuro-evolution methods for gathering and collective construction

Abstract

This paper evaluates the Collective Neuro-Evolution (CONE) method, comparative to a related controller design method, in a simulated multi-robot system. CONE solves collective behavior tasks, and increases task performance via facilitating behavioral specialization. Emergent specialization is guided by genotype and behavioral specialization difference metrics that regulate genotype recombination. CONE is comparatively evaluated with a similar Neuro-Evolution (NE) method in a Gathering and Collective Construction (GACC) task. This task requires a multi-robot system to gather objects of various types and then cooperatively build a structure from the gathered objects. This collective behavior task requires that robots adopt complementary and specialized behaviors in order to solve. Results indicate that CONE is appropriate for evolving collective behaviors for the GACC task, given that this task requires behavioral specialization. © 2011 Springer-Verlag.