Abstract
Swarm robotics is the research field of multi-robot systems which consist of many homogeneous autonomous robots without any type of global controllers. Designing swarm robotic systems (SRS) is difficult since the accurate system model of a robotic swarm is usually not available. In this paper, thus, a reinforcement learning (RL) approach is applied for controlling each robot. RL allows participating robots to learn mapping from their states to their actions by rewards or payoffs obtained through interaction with their environment on the trial and error basis. However, RL in a conventional framework does not perform well because an environment of SRS is dynamic and continuous. We have been developing a technique which segments state and action spaces autonomously to extended the adaptability to a dynamic environment. In order to improve the learning performance of this technique, an ensemble learning mechanism, in which output is combined of multiple outputs, is introduced. We evaluate our proposed method by conducting experiments for a cooperative carrying task with ten autonomous mobile robots.
