Introducing Lifelong Learning in Swarm Robotics

This chapter presents the first known attempt to enable a lifelong machine learning system for adaptive emergent behaviours in swarming agents. The proposed system builds on our decentralised swarm behaviour selector, R-HGN, and uses post-operation simulations to improve behaviour selection and expand the swarm’s behaviour repertoire. Via simulation, we demonstrate that this lifelong learning swarm is able to learn from past experiences, emulating hindsight, and use the collective findings of all swarm members to improve the swarm as a whole, emulating an adaptive culture. The learning swarm is evaluated in a nontrivial task and found to outperform a once-trained swarm by up to 23% of the fitness range.