AGENT-BASED SOCIAL SIMULATION

Engineering Agent-Based Social Simulations (EABSS)

The EABSS framework is a structured, transparent approach that integrates best practices from Software Engineering for ABSS modeling. It supports various activities from model conceptualization to documentation and the identification of novel research directions. Grounded in co-creation principles and drawing on software engineering and service design elements, EABSS provides a step-by-step process for detailed system analysis.

It includes predefined interactive tasks, table templates, and UML diagrams for stimulating and documenting contributions from all participants. The framework's iterative nature ensures that information from previous steps can be reused, acting as a built-in validation mechanism. The outcome is a structured record that is easily understandable by all stakeholders and implementable as a simulation model.

The Role of Conversational AI Systems (CAIS) & LLMs

Conversational AI Systems (CAIS), powered by Natural Language Processing (NLP) and Generation (NLG) models like Large Language Models (LLMs) such as ChatGPT, are designed to engage in human-like conversations. They understand context, discern intent, and generate coherent responses based on vast pre-trained data and machine learning. While powerful, CAIS have limitations, including struggles with ambiguity and a lack of real-world understanding, necessitating their use as supplementary tools.

In ABSS, LLMs show promise in enhancing simulated agent decision-making, social interactions, and communication. Generative ABM (GABM), where agents are connected to LLMs, allows for more nuanced and realistic behaviors than traditional rule-based models. This research specifically explores how CAIS can directly support the conceptual modeling phase of ABSS, addressing a gap in current literature by demonstrating practical applications.

Prompt Engineering for ABSS Model Design

This research adopts a robust methodology, focusing on CAI as an idea generator and an impersonator for absent stakeholders in ABSS model design. Key to this is Prompt Engineering, which currently operates more as an art than a science. The approach utilizes a standard prompt template, emphasizing context, clear instructions, constraints, and feedback loops to ensure high-quality outputs from CAIS.

A comprehensive "EABSS script" was developed, adhering to principles of ease of use, accessibility (ChatGPT 3.5), response quality, and structured output (tables, diagrams). The script is segmented into analysis, design, and concluding remarks, using prompt design strategies like chat preparation, role/tone definition, and information reuse to manage CAIS's stateless nature and memory limitations effectively. Fine-tuning of the LLM is achieved by leveraging the iterative information retrieval of the EABSS framework.

Illustrative Case Study: Adaptive Architecture in a Museum Context

To validate the EABSS script and the utility of CAIS, an illustrative case study on "Adaptive Architecture in a Museum Context" was employed. The goal was to generate innovative conceptual ideas for an ABSS model simulating visitor interactions with dynamic museum environments. This included two primary adaptive architecture artefacts:

• Large wall-mounted screens with smart content windows that move with visitors.
• A smart partition wall that dynamically reconfigures the exhibition space based on real-time visitor movement.

The CAIS, guided by the EABSS script, successfully generated detailed outputs, including problem statements, study outlines, experimental factors, and UML actors, objects, and diagrams. Despite minor inaccuracies, the CAIS proved to be a valuable co-creation companion, offering fresh perspectives and accelerating the design process compared to traditional methods.

Benefits & Future Opportunities of CAIS in ABSS

The integration of CAISs like ChatGPT into ABSS model design offers significant advantages, primarily in efficiency, transforming weeks/days of conceptualization into hours. It enables the creation of multiple focus group scenarios by defining virtual personas and tones, simulating diverse stakeholder perspectives. The EABSS script’s adaptability allows for minimal modifications across various case studies, making it a versatile tool for researchers.

However, limitations exist, including CAIS's struggle with causality comprehension and potential biases from training data, which could affect model accuracy and ethics. Future work should focus on automating prompt submission via APIs, training specialized LLMs for social simulation, and further refining prompt engineering to enhance discussions and diagram generation quality. Despite these challenges, CAIS emerges as a compelling aid for ABSS model developers, poised to significantly influence future model development practices.