← Back to Research

An LLM-Driven Simulation Framework for Environmental Human Behavior

with Y. Yoo & D. Shin
Work in Progress

Abstract

This research develops a novel Person-Environment-Situation (P.E.S.) framework that extends Lewin's behavioral theory for LLM-based spatial choice prediction. By integrating environmental psychology with large language models, we create a comprehensive simulation system that can predict human behavior in various spatial contexts, particularly valuable for Cold-Start scenarios where historical behavioral data is unavailable.

Research Motivation

🧠 Theoretical Foundation

Kurt Lewin's field theory B = f(P,E) suggests that behavior is a function of both person and environment. However, traditional computational models struggle to capture the nuanced interactions between:

  • Individual psychological factors and preferences
  • Dynamic environmental characteristics
  • Situational contexts that modify behavior patterns
  • Real-time adaptation to changing conditions

🚀 Cold-Start Problem

Many spatial prediction systems fail when:

  • New environments lack historical usage data
  • Novel user groups without established behavioral patterns
  • Unique situational contexts not seen in training data
  • Need for immediate deployment without extensive data collection

P.E.S. Framework

👤 Person (P) Component

  • Demographics: Age, gender, cultural background
  • Psychological Traits: Personality, preferences, risk tolerance
  • Goals & Motivations: Current objectives and intentions
  • Past Experiences: Relevant behavioral history and learning

🌍 Environment (E) Component

  • Physical Layout: Spatial configuration and accessibility
  • Social Context: Presence and behavior of others
  • Ambient Conditions: Lighting, noise, temperature, crowding
  • Affordances: Available actions and interaction possibilities

📍 Situation (S) Component

  • Temporal Context: Time of day, season, duration constraints
  • Social Dynamics: Group composition and social pressures
  • Task Context: Primary and secondary objectives
  • External Constraints: Rules, regulations, and limitations

Methodology

🤖 LLM Integration Architecture

Multi-layer prompt engineering approach:

  • Context Encoder: Transforms P.E.S. variables into structured prompts
  • Behavioral Reasoning: LLM processes psychological and environmental factors
  • Choice Prediction: Generates spatial choice probabilities with reasoning
  • Confidence Assessment: Self-evaluation of prediction reliability

📊 Validation Methodology

  • Dataset: 789K real usage records from diverse spatial environments
  • Baseline Comparison: Traditional rule-based and ML approaches
  • Cross-validation: Temporal and spatial generalization testing
  • Cold-Start Evaluation: Performance on completely new environments

🔄 Multi-Agent Implementation

  • Agent Architecture: Individual P.E.S.-driven LLM agents
  • Interaction Modeling: Social influence and collective behavior
  • Dynamic Adaptation: Real-time learning from environmental feedback
  • Scalability Design: Efficient processing for large-scale simulations

Key Findings

📈

83% Accuracy Improvement

P.E.S. framework achieved 83% improvement over existing methods in spatial choice prediction accuracy.

🎯

Cold-Start Excellence

Framework maintains 78% accuracy even in completely new environments without historical data.

🧠

Reasoning Transparency

LLM provides interpretable reasoning for each prediction, enabling human understanding and validation.

⚡

Real-time Performance

Multi-agent system processes 1000+ simultaneous behavioral predictions with sub-second response times.

Validation Results

📊 Prediction Accuracy

P.E.S. Framework: 89.2%
Traditional ML: 48.7%
Rule-based Systems: 35.1%

🆕 Cold-Start Performance

New Environments: 78.3%
New User Groups: 82.1%
Novel Situations: 74.9%

⏱️ System Performance

Response Time: 0.8s
Concurrent Users: 1,000+
Reasoning Quality: 91.5%

Applications & Impact

🏢 Smart Building Management

Predict occupancy patterns and optimize space utilization in real-time for new building layouts.

🎯 Personalized Recommendations

Provide location-based suggestions even for first-time visitors without behavioral history.

🚨 Emergency Response

Model evacuation behavior in crisis situations across diverse populations and environments.

📊 Urban Planning

Simulate human behavior in proposed urban developments before construction begins.

Future Directions

🔮 Research Extensions

  • Integration with IoT sensors for real-time environmental updates
  • Cross-cultural validation across different geographical regions
  • Temporal behavior modeling for long-term pattern prediction
  • Integration with VR/AR for immersive behavior simulation

🛠️ Technical Improvements

  • Fine-tuning specialized LLMs for spatial reasoning tasks
  • Development of lightweight models for edge computing
  • Enhanced multi-modal input processing (visual, audio, sensor data)
  • Federated learning for privacy-preserving behavior modeling

Technical Specifications

Large Language Models Environmental Psychology Multi-Agent Systems Spatial Analytics Behavioral Modeling Cold-Start Prediction Real-time Processing Human-Computer Interaction