A Study on Online Hyper-heuristic Learning for Swarm Robots

Swarm robots continue to become more prominent in solving challenging tasks in real world applications. Due to the complexity of operating in often unknown environments, centralised control of swarm robots is not ideal. Prior manual programming is also not practical under these kind of circumstances. Thus, we establish a hyper-heuristic based learning approach for swarm robot control. With this framework, robots can autonomously identify appropriate heuristics from a set of given low-level heuristics, each heuristic guiding certain behaviours. We evaluated this type of online learning on building surface cleaning and studied the effectiveness of our hyper-heuristic online learning. Nine heuristics were proposed in this study. Through the experiments it can be seen that robots can improve their cleaning performance through the online learning process. More importantly, the experiments show that appropriate heuristics can be selected even when the size of the heuristic set is changed. The study on four types of environments shows that with the same heuristic set, the robot swarm can adapt to different environments for different tasks. Hence, hyper-heuristic learning is an effective method for decentralised control of swarm robots.