Towards City Simulation Using LLM Agents

Urban planning, social science, and behavioral studies have long sought realistic simulations of human behavior within urban environments. Traditional agent-based simulations, however, often fall short, relying on rigid, handcrafted rules that can't capture the complexities of human intentions and adaptive behaviors.

CitySim emerges as a groundbreaking framework, integrating large language models (LLMs) to simulate realistic, adaptive urban behavior on an unprecedented scale.

What is CitySim?

CitySim leverages advances in large language models (like GPT-4o) to create agents that act and interact realistically within a virtual urban environment. These agents are equipped with:

• Diverse Personas: Each agent has demographic attributes, personality traits, and preferences derived from real-world surveys.
• Recursive Value-Driven Planning: Agents dynamically generate daily schedules considering mandatory tasks, personal habits, situational factors, and intrinsic desires.
• Long-Term Goals and Memory: Agents maintain evolving beliefs, long-term aspirations, and memories that affect their decisions and interactions over time.

Key Innovations

Cognitive State Representation

Persona Module

CitySim agents are initialized with detailed personas derived from survey data, including demographic attributes (age, occupation, education), personality traits (Big Five), and habitual behaviors (e.g., meal timings, leisure activities).

Memory Module

Agents maintain:

• Temporal Memory: Chronological records of daily events.
• Reflective Memory: Agents synthesize daily events into higher-level insights.
• Spatial Memory: Beliefs about visited locations, updated through direct observations and similarity-based inference.

Belief Module

Upon visiting a location, agents form subjective appraisals based on their persona and context, continuously refining their beliefs about different city locations.

Mobility Behaviors

Agents make sophisticated decisions regarding their daily movements:

• Recursive Daily Planning: Filling schedules starting from mandatory activities, down to leisure activities based on their intrinsic values.
• Place Selection: A belief-weighted gravity model guides location choices, balancing personal preferences and proximity.
• Vehicle Selection: Transportation mode is chosen by evaluating factors like distance, time, weather, and individual preferences.

Social Behaviors

CitySim agents form dynamic social relationships, maintaining evolving beliefs about affinity, trust, and familiarity with others, leading to realistic face-to-face and online interactions.

Experimental Results

CitySim significantly outperforms traditional and other LLM-based simulation frameworks in multiple dimensions:

Macro-level Time Use

Agents' daily schedules closely match real-world survey data, demonstrating realistic macro-level patterns of human activities.

Behavioral Realism

CitySim agents were consistently rated as more human-like than agents from other leading frameworks, thanks to adaptive, context-sensitive behaviors.

Travel Patterns

Simulated travel data from CitySim closely aligns with real-world patterns, accurately reflecting peaks and troughs in urban mobility.

Crowd Density

CitySim effectively models pedestrian crowd density, closely matching real-world data in major urban areas.

Limitations and Ethics

Despite its strengths, CitySim inherits potential biases from underlying LLMs and may occasionally produce inaccuracies. Ethical considerations include the risk of amplifying biases or influencing real-world urban policies without adequate human oversight.

Conclusion

CitySim represents a significant advancement towards realistic city-scale simulation, enabling nuanced insights into human urban behaviors, beneficial for research, urban planning, and policy-making.

CitySim sets a new standard in agent-based modeling, moving beyond rigid rules and embracing the adaptability and complexity of human behaviors.

Link to the original paper:
CitySim: Modeling Urban Behaviors and City Dynamics with Large-Scale LLM-Driven Agent Simulation