A scheme proposal for the development of machine learning-driven agent-based models through case studies

Abstract

Agent-based modeling (ABM) has garnered extensive application across various disciplines, presenting an instrumental tool for researchers. However, the creation of effective and precise agent-based models (ABMs) presents an intricate challenge. The integration of machine learning (ML) methodologies into ABMs could potentially simplify the modeling process and augment model performance. This dissertation presents an exhaustive exploration of the relationship between ABM and ML methodologies, elucidating the advantages and disadvantages of data-driven ABMs. The predominant challenge in the development of ABMs is the delineation of agents' behavioral rules. To surmount this challenge, a main scheme utilizing ML methodologies within ABMs is proposed. This scheme is examined through the lens of pertinent academic literature. Within this central framework, three sub-frameworks are introduced to model agent behaviors in ABMs via ML methodologies. Each of these sub-frameworks pivots around the principle of employing ML methodologies to derive agent behaviors. Sub-Framework I offers a comprehensive roadmap for researchers aiming to implement an ABM with predictive capabilities. This sub-framework intertwines theoretical support with ML methodologies with the aim of enhancing the precision of ML-generated agent behavior. To determine how well ML techniques enhance the accuracy and ease of building ABM, Sub-Framework I was applied to a real-world case using supervised learning methods. Sub-Framework II offers guidance to researchers with the objective of creating ABMs for optimization purposes. The effectiveness of this framework is scrutinized in a simulated environment employing Reinforcement Learning (RL) techniques as the ML methodology. Lastly, Sub-Framework III serves as a guide for researchers endeavoring to create an ABM for understanding objectives. Its empirical validation is undertaken through a real-world case study. This framework utilizes Inverse Reinforcement Learning (IRL) and Reinforcement Learning (RL) as ML methodologies. The findings of these models developed through the frameworks suggest that ML approaches may facilitate the development of ABMs.