Research theme

Agent-Based Modeling, Human Mobility, and LLM-based Analysis

We use agent-based models to connect hazard conditions and urban environments with the decisions, movements, and interactions that shape disaster response (e.g., evacuation, resource allocation, and decision-making).

← Research overviewTUNAMI-EVAC on GitHub ↗ GTEM on GitHub ↗

Why ABM?

From hazard maps to human response

Physical models estimate hazard intensity and remote sensing identifies impacts. Agent-based models add the human layer: who evacuates, when they start, which route they choose, how congestion develops, and how interventions change the result.

Our work is aimed at both scientific understanding and practical use. The Global Tsunami Evacuation Model (GTEM) develops a simple, open and documented workflow for local governments, while other projects (e.g., TUNAMI-EVAC) explore behavioral realism, adaptive guidance, and Digital Twin integration.

Selected work

From open models to adaptive guidance

Global Tsunami Evacuation Model

GTEM builds on earlier NetLogo evacuation research and moves toward a Python-driven, practitioner-oriented workflow.

  • Open and adaptable case-study structure
  • Scenario testing for roads, population, departure behavior, and safe zones
  • Automated reports, maps, congestion diagnostics, and optional animations
  • Manuals that explain what the model can and cannot answer

TUNAMI-EVAC

Maintenance and development of an open tsunami evacuation model that represents pedestrians moving through urban road networks.

  • Agent-based pedestrian movement
  • GIS/network-based evacuation
  • Population and age-group inputs
  • Reusable foundation for future scenarios and teaching

Behavior and social interaction

Our research explores how stress, crowd conditions, social-force concepts, and heterogeneous response affect evacuation dynamics.

  • Delayed departure and reaction distributions
  • Age-dependent movement
  • Congestion and density effects
  • Future extensions for vehicles and mixed traffic

Reinforcement-learning guidance

In the 2024 study, node-agents at intersections learn route recommendations based on local congestion rather than sending everyone along the shortest path.

  • SARSA reinforcement learning
  • Density-dependent speed
  • Adaptive intersection-by-intersection guidance
  • Strongest benefits during early high-density evacuation
Read the paper ↗

Disaster Digital Twins

Connects evacuation simulation with 3D urban data, people-flow information, infrastructure, sensors, and decision-support layers.

  • Static and dynamic data integration
  • Normal-time and disaster-time use cases
  • Scenario comparison for bottlenecks and vulnerabilities
  • Future adaptive guidance through signs, signals, or mobile systems

Co-design with practitioners

Through the CDRI Fellowship, local-government feedback informs the functions, outputs, and documentation of the model.

  • Questionnaires and stakeholder engagement
  • Case studies for coastal communities
  • Simple inputs and interpretable outputs
  • Training and dissemination through international partners
OpenModels, documentation, and case studies designed for reuse.
Multi-scaleFrom individual movement to city-scale evacuation networks.
TestableScenario comparison, stochastic runs, and convergence checks.
ActionableMaps, curves, reports, videos, and planning questions.

Practical outputs

What can an ABM help answer?

  • How long does evacuation take under different population and warning scenarios?
  • Where do bottlenecks form, and which streets are most critical?
  • How are people distributed across safe zones or evacuation buildings?
  • How do road disruptions, new shelters, or guidance strategies change outcomes?
  • Which groups or areas remain vulnerable when time is limited?