Robotic Hierarchical Graph Neurons. A novel implementation of HGN for swarm robotic behaviour control

• Published in: Expert systems with applications Vol. 186; p. 115675
• Main Authors: Smith, Phillip, Aleti, Aldeida, Lee, Vincent C.S., Hunjet, Robert, Khan, Asad
• Format: Journal Article
• Language: English
• Published: New York Elsevier Ltd 30.12.2021; Elsevier BV
• Subjects: Behavior; Distorted pattern recognition; Hierarchical Graph Neurons; Neurons; Pattern matching; Robot control; Robotic behaviour; Robotics; Robots; Swarm robotics; Swarming; Distorted pattern recognition; Hierarchical Graph Neurons; Env; Robotic behaviour; Swarm robotics
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2021.115675

Simple rule-based robot behaviours, such as those utilised for swarming robots, typically excel in only the niche conditions for which they were designed. Behaviour selection allows robots to switch between these specialised behaviours in accordance with the observed conditions. This paper explores the use of a novel form of Hierarchical Graph Neurons (HGN) for such behaviour selection within a swarm of robotic agents. This new HGN is called Robotic-HGN (R-HGN) as it allows pattern matching of mixed datasets of robot observations. R-HGN matches said patterns to labelled environments and allows appropriate robot behaviours to be utilised throughout an operation in a ‘society of mind’ approach to task flexibility in robots. This approach is novel to the HGN field as it expands the application beyond discrete categorical data inputs. Additionally, this research is novel to the field of robotic swarming as it explores a new method to temporal agent diversity for overcoming localised environment challenges. This R-HGN for behaviour selection is validated against individual behaviour implementations and a random behaviour selection. The comparison is made via statistical distribution of swarm fitnesses in multiple instances of a non-trivial swarming task. From this comparison R-HGN is found to enable appropriate behaviour selection in both environments known and unknown a priori, resulting in a median swarm performance improvement of up to 389%. Finally, in environments prior observed, the R-HGN environment prediction one-versus-all accuracy is up to 99.1% and F1 scores reach a maximum of 97.15%. •Hierarchical graph neurons can be altered for robot observation patterns.•Robotics can use patterns to identify environment and select optimal behaviour.•Environment matched in 92.4-98.5% of pre-seen cases.•Optimal behaviour utilised in 78% of environments prior unseen.