Heterogeneity in Multi-Agent Systems

Abstract

Heterogeneous multi-agent systems are composed of diverse and autonomous agents that can interact and cooperate to achieve complex goals. Human history from the formation of social groups to technological systems such as the Internet has shown that greater functionality is achieved in interconnected systems. It is therefore expected that advances in artificial intelligence and autonomy will lead to diverse types of ever more capable robotic and software agents interacting in order to enhance their capabilities.

In applications including logistics, disaster relief and social care, these heterogeneous agents can bring different perspectives, skills, and resources to a system, enhancing its adaptability, robustness, and creativity. And yet, despite the benefits, it is not clear what constitutes heterogeneity in this context or how to frame it as a property in system design. Without answering these questions, there is a risk that the full benefits of multi-agent systems are not realised, and that their collective behaviour either surprises or at worst is detrimental to those it is intended to serve.

This thesis addresses the need to understand what heterogeneity means in the context of a multi-agent system, and the tools and techniques to employ different agent types effectively in system design. It explores how the context, task and agent interactions, in conjunction with the number of agent types and their distribution can influence how heterogeneity is framed and the methods used to measure it.

The research also investigates general design principles to facilitate knowledge transfer between applications, thereby reducing development time and the risk of failures. Application agnostic techniques based on the information-theoretic measure of Empowerment and Evolution are investigated to create successful interactions between agents regardless of their type. Additionally, the concepts of influence, traits, and ecological framing of diversity are explored for their relevance to artificial systems.

By gaining a more comprehensive understanding of heterogeneity in multi-agent systems, this research contributes to the development of systems that capitalise on the benefits of heterogeneous agents while minimizing potential negative consequences arising from the presence of mixed agent types.

Date of Award	7 May 2024
Original language	English
Awarding Institution	University of Bristol University of the West of England
Supervisor	Jonathan Lawry (Supervisor), Seth Bullock (Supervisor), David Harvey (Supervisor) & Angus Johnson (Supervisor)