The Emergence of Complex Behaviours in Agent-Based Models using Reinforcement Learning

Over the last two decades, agent-based modelling (ABM) has become an invaluable tool across various disciplines for simulating the intricacies of complex systems. At the core of these models are the decision-making frameworks that guide agent behaviour. Traditionally, these have been constrained to predetermined rules, providing deterministic insights yet often omitting the dynamic process of learning and adaptation inherent in real-world behaviour. This research examines the integration of Reinforcement Learning (RL) within a set of ABMs—designated as 'hybrid-ABMs'. These models are characterised by their ability to merge traditional ABMs with the adaptive, experience-driven learning processes facilitated by RL. By doing so, this thesis explores to what extent such integration enables agents to develop and refine intelligent behaviours in response to environmental stimuli and changing scenarios. Focusing on three distinct domains—predator-prey ecosystems, burglary and criminal behaviour, and the economic behaviours of housing markets—this thesis assesses how RL can empower ABMs to simulate complex phenomena with greater detail. Through these hybrid-ABMs, we scrutinise whether RL-enhanced agents can autonomously acquire behaviours that not only resonate with theoretical and empirical findings but also adapt to unforeseen circumstances with a degree of experience and reliability. The findings are: agents within these hybrid-ABMs successfully learn from their environment, the emergent behaviours align with established literature in the respective fields, and the agents exhibit a capacity to adapt to novel situations effectively. This thesis hypothesises that neurologically inspired algorithms like RL can enhance sociological ABMs by introducing an element of learning and adaptability, thereby equipping these models to better mirror the complexity of real-world systems and decision-making processes.