ºÃÉ«ÏÈÉúTV

Development of a Disaster Simulation Research & Training Platform

AI-SDM is partnering with the American Red Cross to create a high-fidelity designed to address the critical challenges of disaster management training. This platform serves as both a sophisticated research environment and a pedagogical tool, focusing on the intricate trade-offs inherent in distributing limited resources—such as emergency shelter, food supplies, and specialized personnel—under conditions of extreme temporal and spatial uncertainty. By integrating interactive human game engines with decentralized multi-agent reinforcement learning (MARL) and large language models (LLMs), the game supports multiple operational modes within human–AI collaborative configurations. This architecture moves beyond traditional static tabletop exercises by introducing probabilistic event dynamics, including infrastructure failures and evolving flood zones, which require real-time adaptation and strategic flexibility.

Screenshot of the user interface for the disaster resource allocation simulation platform

The technical foundation of the project is a modular, mechanism-first architecture designed for extensibility and experimental control. Rather than relying on linear scripting, the system employs a game database and a grid-based spatial model that supports hazard propagation, resource movement, and evacuation flow through core systemic interactions. New task templates, event triggers, and scenario conditions can be defined declaratively without modifying core code, enabling rapid configuration and controlled variation across experiments. Participants operate under dynamically evolving conditions—including cascading floods and infrastructure failure chains—which enhance ecological realism while preserving analytical tractability. The platform logs fine-grained state transitions and decision traces, allowing researchers to examine decision strategies, trust calibration, and the influence of AI-guided interventions in high-pressure environments.

Example of an emergency event in the game (damaged Emergency Response Vehicle)Beyond the immediate training applications, the game generates granular behavioral trace data, providing a rich dataset for cognitive modeling. Using frameworks like Instance-Based Learning (IBL), the institute can analyze decision trajectories to understand how experts and novices navigate multiple competing tradeoffs, including speed–accuracy (rapid deployment vs. better information), risk–ambiguity (known vs. uncertain needs), and equity–efficiency (serving the hardest-hit vs. the most accessible communities). The platform’s evaluation framework tracks multi-dimensional metrics, including efficiency and the downstream effects of delayed responses. Through a chat-based dialogue system, AI agents offer explainable suggestions, allowing researchers to investigate the calibration of trust and the efficacy of model-based guidance in high-pressure environments. Moving forward, AI-SDM plans to integrate more advanced cognitive models into the agents to replicate human-like reasoning flaws, such as anchoring, further refining our understanding of how AI can effectively steer human behavior toward optimal outcomes during societal crises. What’s more, by systematically manipulating AI agents’ recommendations, researchers can examine how humans interpret, use, and rely on AI-provided information.