AI-Enhanced ABM Development: Facilitating Agent-based Modeling using Artificial Intelligence

Abstract

Agent-based Modeling (ABM) is a multi-disciplinary micro-to-macro level simulation technique that is based on autonomous entity interaction – entities with the ability to learn, make decisions, and display random behaviors – in heterogeneous systems (a mechanism of systems that have their independent mechanisms while functioning together in equilibrium) which aims to help us in understanding complexities of the real world, decision, and policy-making. The emergence of ABM is following a scientific paradigm shift in the methodologies of studying social systems. Traditional methods and techniques from various fields like philosophy, mathematics, and prior simulation techniques such as Discrete Event Simulation (DES) and System Dynamics (SD) have been insufficient in describing the aggregated outcomes of interactions among autonomous entities. Modeling the spread of COVID-19 in an environment, the interaction of suppliers and consumers in the market, the mechanism of studying traffic, and the population growth scenarios are some examples of simulations to which ABM has contributed. ABM has immense capabilities in capturing complex details; however, despite its rich capabilities, it demands a complex set of skills to cover the underlined prerequests of ABM development, like the design of concepts and implementation of the simulation model. The diverse skill set required by ABM, the collaboration of multiple theoreticians and programmers to support these skills, as well as different sets of qualitative and quantitative data and analysis types, result in many errors, artifacts, and challenges in the process of model development. Comprehensive analysis of narratives, omitted theories and evidence, inconsistent and unstructured documentation, inconsistent simulation coding, nonlinear empirical behavior modeling, data inconsistency, model summary and visualization, validation and verification errors, and model re-usability are areas of challenges associated with ABM development. Although the social simulation community has attempted to address many of these issues through various solutions, including theoretical discussions, standards, frameworks, and protocols, these approaches have inherent limitations. Central limitations build on the conventional perspective of a human-centered development process of ABM, while the emergence and adoption of novel and state-of-the-art computational and Artificial Intelligence (AI) techniques (like Machine Learning (ML), Deep Learning (DL), and Natural Language Processing (NLP)) bear great promises in addressing typical ABM development challenges. This traditional human-oriented perspective highlights the absence of a scientific paradigm to comprehensively address and incorporate AI advancements into the ABM development process. This absence of a revised perspective of ABM in light of recent developments invites a new set of theoretical and methodological discussions to work toward a comprehensive scientific methodology best geared to identify and facilitate ABM development challenges. In this thesis, our aim is to facilitate ABM using AI methods, specifically NLP and ML, to offer a new perspective of ABM development which incorporates a procedural and analytical (qualitative and quantitative) perspective to cover the systematic integrating of these computational techniques. In this context, the first set of contributions in this thesis involves establishing a novel theoretical rationale (Artificial Intelligence Enhanced Agent-based Modeling Development (AIeABM)) to ground the discussions of the new scientific paradigm by incorporating three components. AIeABM initiates the discussion about how to integrate AI schema to establish a bigger picture of ABM development steps and their relevant data types in order to discuss and select the appropriate computational techniques. To this end, we have proposed a synthesized ABM life cycle (Synthesized Development Process for Agent-based Modeling (SDP4ABM)). This framework aims to provide a structured approach to discuss possible AI techniques and categorizing challenges, both from procedural and analytical perspectives, to effectively utilize novel and state-of-the-art AI techniques. Given SDP4ABM as the first component and a comprehensive analysis of AI techniques from both application-based and modeling perspectives as the second component, AIeABM serves as the third component, bringing together the overall picture of ABM development challenges that can be effectively addressed through AI integration. This structured approach ensures a cohesive understanding of how AI can facilitate various stages of the ABM life cycle, from problem conceptualization to model implementation and beyond. Building on the SDP4ABM schema and focusing on the transition from conceptualization and implementation stages of ABM development, we have proposed methods and techniques to address both qualitative and quantitative challenges. As our second contribution, we have introduced a novel NLP-based methodology and a corresponding benchmark utilizing Question-answering (QA) models to extract, summarize, and represent information from conceptual models. This approach is designed to be used by both human users and computational Large Language Model (LLM) to facilitate automated code generation. While the proposed benchmark provides a foundation for future developments and comparative studies, the methodology itself can be extended to other stages of SDP4ABM to address additional qualitative challenges, such as narrative analysis. As the third contribution, and turning to the operationalization of quantitative data, we have developed a Python package named ‘A Correlation Pattern Recognizer Python Package for Nonlinear Relations (Copatrec)’, which leverages empirical optimization techniques from ML to enhance the mathematical representation of behaviors in ABM development using empirical data. This package automates processes such as data pre-processing, nonlinear model selection, model validation, and verification. To demonstrate the tool’s efficiency and robustness, we conducted two case studies in ABM behavior development of poverty and statistical nonlinear analysis of the same phenomena. One essential learning of these contributions is the indication of a paradigm shift in the social simulation community in response to commonly observed ABM development challenges, necessitating the consideration of novel computational techniques like NLP and ML in the process of ABM development. This thesis is not the first work utilizing AI for ABM development. Since the recent acceleration in the development of AI and ML techniques, various attempts have been made to address selected challenges. However, it is essential to have a broader understanding of the involved process in order to systematically integrate and contextualize future contributions. While the novelty, robustness, and consistency of these contributions have been discussed in their respective articles and sections, there remain opportunities for future development. This includes further development of the AIeABM paradigm, the SDP4ABM schema discussion, technical aspects such as the expansion of the NLP benchmarks and techniques to address additional stages of ABM challenges, as well as enhancing our Copatrec package. Moving beyond the immediate contributions, the thesis concludes with a discussion of the general implications of AIeABM for the area of agent-based modeling.