AIDRA: An Intelligent Disaster Rescue System Integrating Six AI Technologies

Disaster rescue is a race against time for lives. Traditional rescue relies on human experience and struggles to process massive dynamic information in real time. The AIDRA (Adaptive Intelligent Disaster Response Assistant) project proposes a hybrid AI architecture that integrates search algorithms, constraint satisfaction problems (CSP), machine learning (ML), reinforcement learning (RL), and fuzzy logic, among other technologies, to build an intelligent rescue system with autonomous decision-making and real-time adaptation, aiming to address the uncertainty challenges at disaster sites.

Disaster sites are full of uncertainties (fire spread, road blockages, worsening of injured people's conditions, aftershocks, etc.). Single AI technologies have obvious shortcomings: search algorithms are good at path planning but cannot handle ambiguous danger levels; machine learning can predict survival probabilities but struggles to explain decision logic; reinforcement learning can adapt to the environment but requires a lot of trial and error. Therefore, AIDRA adopts a layered architecture, allowing different AI paradigms to work together to form a complete decision loop.

Seven-Layer Architecture: User layer (interaction interface), presentation layer (Pygame/Tkinter real-time visualization), simulation layer (grid state/hazard area/injured health management), search and planning layer (path calculation), AI decision layer (ML/RL/fuzzy logic integration), constraint satisfaction layer (resource allocation), analysis and data layer (performance monitoring/decision logs).

Core Technologies:

• Intelligent path planning: Optimized A* algorithm with fire penalty weight, reducing path length in dangerous scenarios by 12-18%;
• Resource allocation: CSP solver (backtracking + MRV heuristic + forward checking) reduces search nodes by 65% and allocation delay by 50%;
• Survival prediction: KNN+MLP hybrid model achieves 91% accuracy;
• Dynamic learning: Q-Learning agent based on Epsilon-Greedy strategy for trial-and-error learning;
• Fuzzy logic: Handles ambiguous information like "dangerous" and "urgent" to calculate continuous scores.

Simulation Environment: A 10×11 2D grid that includes dynamic fire areas, random road blockages, random aftershocks, and health decay of injured people, simulating real disaster pressure.

Transparent AI: Provides a real-time visualization dashboard, including real-time confusion matrix (prediction performance), algorithm battle (decision difference comparison), and side-by-side path comparison (decision basis display), enhancing commanders' understanding and trust in AI decisions.

Practical Effects: Navigation efficiency improved by 12-18%, survival prediction accuracy of 91%, resource allocation delay reduced by 50%.

Development Background: Developed by Maryam Khalid and Rohan Munir from the Islamabad Campus of Bahria University.

Significance: Demonstrates the value of hybrid AI architecture in high-risk fields, providing reference for complex scenarios such as medical first aid and industrial safety.

Future expansion directions for AIDRA: Access to real GIS map data, integration of drones and IoT sensors, multi-agent collaborative rescue, and connection to real emergency systems. With the development of edge computing and 5G technology, it is expected to become a life-saving tool in actual combat.