Generating and Analyzing Collective Behavior in a Robotic Swarm by the Use of Deep Reinforcement Learning and Deep Neuroevolution

Abstract

This study proposes a method to apply deep neural networks to controllers of robotic swarms. In a typical approach to design controllers, the designer has to define the features extracted from sensory inputs in advance. By applying deep neural networks with convolution layers, it can automatically extract features from sensory inputs. We applied two methods to train the deep neural networks, i.e.,deep reinforcement learning and deep neuroevolution. The controllers were tested in a path-formation task using computer simulations. Compared with deep reinforcement learning, deep neuroevolution was able to generate collective behavior even in sparse reward settings.