Leveraging Heterogeneous Controller Representations for Evolutionary Swarm Robotics

Abstract

A dominant approach to developing multi-agent swarm systems is the automatic design of individual agent controllers via evolutionary computation. Typically, these controllers are encoded as neural networks. However, controllers can be represented in many other ways, and recent work suggests that different representations excel at different tasks in single agent systems. In particular, Cartesian genetic programming and Markov brains are two promising representations that share neural networks' ability to be applied to generic problems without extensive task-specific experimenter input. Here, we extend prior results to agents within swarms, and show that there likely exist swarm tasks for which neural networks are not the best suited controller representation. Moreover, many swarm problems have subcomponents that would benefit from division of labor among specialist agents, making them amenable to a modular design. Given the relatively low probability that all subproblems happen to be optimally solved by the same controller type, swarms composed of multiple controller types may outperform swarms which have access to only one type of controller representation. We show that such heterogeneous swarms consistently specialize and achieve high performance by evolving to use controller representations on the subtasks for which they are best suited. Finally, we introduce a simple method to evolve the number of agents using each controller representation in a swarm. These evolved compositions perform well when evaluated against both homogeneous swarms and swarms with preset compositions.