Bin Kamruddin, Ayman (2024) Collective decision making in multi-agent systems with applications to human-AI interaction. [Tesi di dottorato]

Documento PDF Final_Thesis_PhD_MERC_XXXVIciclo_binKamruddin_Ayman.pdf Visibile a [TBR] Amministratori dell'archivio Download (12MB) | Richiedi una copia

Abstract

The research presented in this thesis aimed to model human behaviour in complex perceptual-motor tasks. It starts by investigating whether dynamical perceptual-motor primitives (DPMPs) could also be used to capture human navigation in a first-person herding task. To achieve this aim, human participants played a first-person herding game, in which they were required to corral virtual cows, called target agents, into a specified containment zone. In addition to recording and modelling participants' movement trajectories during gameplay, participants' target selection decisions (i.e., the order in which participants corralled targets) were recorded and modelled. The results revealed that a simple DPMP navigation model could effectively reproduce the movement trajectories of participants and that almost 80\% of the participant's target selection decisions could be captured by a simple heuristic policy. Importantly, when this policy was coupled to the DPMP navigation model, the resulting system could successfully simulate and predict the behavioural dynamics (movement trajectories and target selection decisions) of participants in novel multi-target contexts. Building on these findings, the study further explored multi-herder and multi-agent herding scenarios, which introduce additional layers of complexity. In these scenarios, human participants coordinated their movements with one another, dynamically adjusting their strategies based on the behaviour of both their fellow herder and the target agents. Modelling results revealed that human herding behaviour in these multi-agent contexts could be replicated using simple control rules and decision-making processes. Artificial agents were then developed - using these models and additionally, Deep Reinforcement Learning techniques - in order to cooperate with human agents to complete the multiagent herding ask. By integrating multi-herder dynamics into the model, the study provides insights into how artificial agents could be designed to work alongside humans in tasks requiring collaborative joint action, such as search and rescue operations, crowd management, or autonomous driving systems, where multiple entities must cooperate to achieve shared goals.

Downloads

Actions (login required)

Modifica documento