Evolving Test Environments to Identify Faults in Swarm Robotics Algorithms

Swarm robotic systems are often considered to be dependable. However, there is little empirical evidence or theoretical analysis showing that dependability is an inherent property of all swarm robotic systems. Recent literature has identified potential issues with respect to dependability within certain types of swarm robotic control algorithms. However, there is little research on the testing of swarm robotic systems; this provides the motivation for developing a novel testing method for swarm robotic systems. An evolutionary testing method is proposed in this thesis to identify unintended behaviours during the execution of swarm robotic systems autonomously. Three case studies are carried out on flocking control algorithm, foraging algorithm, and task partitioning algorithm. These case studies not only show that the evolutionary testing method has the ability to identify faults in swarm robotic system, but also show that this evolutionary testing method is able to reveal failures in various swarm control algorithms. The experimental results show that the evolutionary testing method can lead to worse swarm performance and reveal more failures than the random testing method within the same number of computing evaluations. Moreover, the case study of flocking control algorithm also shows that the evolutionary testing method covers more failure types than the random testing method. In all three case studies, the dependability of each swarm robotic system has been improved by tackling the faults identified during the testing phase. Consequently, the evolutionary testing method has the potential to be used to help the developers of swarm robotic systems to design and calibrate the swarm control algorithms thereby assuring the dependability of swarm robotic systems.