Slumbers, Oliver; (2025) Population Dynamics in Multi-Agent Systems. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Thesis-4.pdf - Accepted Version Download (9MB) | Preview

Abstract

This thesis explores how population-based approaches can be used to model and solve multi-agent systems (MAS). Traditional game-theoretic methods often struggle with the complexity of large-scale MAS, particularly in adapting to emergent strategic behaviours and computational intractability. Population-based methods offer a scalable alternative by constructing and analysing agent populations (rather than focusing on individuals) that can adapt to evolving dynamics. I investigate this from two perspectives: (1) developing frameworks that iteratively construct populations of agents to enhance MAS learning and robustness, and (2) leveraging generative AI, specifically Large Language Models (LLMs), as a tool for analysing human-like social dynamics in sub-populations. For the first perspective, I focus on three key challenges. First, I examine the role of population diversity in ensuring robustness to a wide range of strategies and in breaking strategic cycles that arise in evolving populations. Second, I address a limitation of traditional opponent selection when building populations: its inability to adapt to the underlying game dynamics by taking a one-size-fits-all approach. Instead, I propose opponent selection approaches that leverage data-driven techniques to dynamically shape population evolution towards finding an optimal population. Third, I challenge the appropriateness of a Nash Equilibrium optimised population in risk-sensitive multi-agent settings, proposing alternative equilibrium concepts and optimisation processes that better integrate risk preferences into population-based training. In the second perspective, I explore the role of language models as agents in social dilemmas, assessing their viability as proxies for human decision-making in multi-agent interactions. I analyse their ability to capture strategic reasoning and emergent behaviours and critically evaluate their offerings in modelling complex population dynamics. This analysis provides insights into the capabilities, and limitations, of using generative AI to study human-like behaviour in MAS sub-populations.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as XMLJSONCSV

Archive Staff Only

View Item