Identifying and modelling decision making and collective behaviour in multi-agent human and artificial systems

The emergence of coordinated action between interacting individuals or agents is a common characteristic of everyday behaviour. Pivotal to the organisation of multiagent activity is the ability of agents to effectively decide how and when to act, with robust decision-making often differentiating expert from non-expert performance. In this thesis we investigated and modelled the behavioural coordination and decision-making behaviour of human and artificial agents completing various herding tasks. Herding tasks involve the interaction of two sets of autonomous agents – one or more herder agents are required to corral a set of heterogeneous target agents. Such activities are ubiquitous in daily life and provide a prototypical example of everyday multi-agent behaviour. We first propose a simple set of local control rules and target selection strategies that enable herder agents to collect and contain a herd of non-cooperative, non-flocking target agents. We then investigated the robustness of the proposed control process to variations in herd size and the strength of the repulsive force that herders imposed on targets. The effectiveness of the proposed approach was also confirmed via ROS simulations and experiments using real robots. We then employed supervised machine learning (SML) to predict the target selection decisions of human herders. The findings demonstrated that the decision-making behaviour of human actors can be effectively predicted using SML at both short (< 1 s) and long (> 10 s) timescales, and that the resultant models can be employed to endow artificial herders with "human-like" decision making capabilities. Finally, we employed explainable AI to understand the state information employed by human herders when making target selection decisions. The findings revealed differences in how expert and novice herders weight state information when making decisions and is the first study to highlight the potential utility of explainable AI techniques for understanding human decision-marking behaviour during multi-agent fast paced interactions.