Hybrid Quantum Machine Learning for Cars: Navigating Natural Disasters Safely

By Dr. Sampath LonkaReviewed by Susha Cheriyedath, M.Sc.Aug 2 2023

In a recent paper submitted to the arXiv* server, researchers investigated the implementation of supervised hybrid quantum machine learning to improve emergency evacuation strategies for cars in the event of natural disasters, particularly earthquakes. The approach involves modeling the problem as a dynamic graph, considering uncertain and ever-changing conditions. The proposed hybrid approach combines a quantum feature-wise linear modulation (FiLM) neural network with a classical FiLM network, mimicking Dijkstra's algorithm.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.

Background

Natural disasters such as earthquakes can lead to catastrophic consequences, including loss of life and property damage. To mitigate these impacts, effective emergency evacuation procedures are crucial, particularly when utilizing cars as a primary mode of transportation. However, earthquakes can significantly disrupt standard road networks due to land deformation, collapsing buildings, and debris, making route planning challenging.

While effective on static graphs, conventional methods such as Dijkstra's algorithm struggle to find optimal paths in dynamically changing situations. This necessitates adapting the algorithm for evolving networks, referred to as the node-wise Dijkstra's algorithm. Nevertheless, obtaining real-time, accurate traffic information is challenging, leading to the need for local information-based solutions that can handle unreliable data.

In the present study, researchers propose a hybrid quantum machine learning approach that leverages quantum technologies to optimize emergency escape plans for cars during natural disasters. Combining classical machine learning and quantum computing, hybrid quantum machine learning techniques have shown promise in addressing industrial-scale problems. The approach is demonstrated through supervised learning (SL), training on node-wise Dijkstra's algorithm decisions in a simulated earthquake scenario with evolving traffic conditions.

Emergency escape routing

The research focuses on developing an effective strategy for rapid city evacuation during emergencies such as earthquakes, floods, or fires, utilizing cars as the means of transportation. The objective is to minimize travel time and bypass traffic congestion while evacuating the city.

The problem of emergency escape routing is designed for a specific map with predefined exit locations and an area affected by an earthquake. As time progresses, traffic near the exit points intensifies, leading to longer travel durations in that vicinity. Additionally, the earthquake continuously impacts roads, resulting in further delays for nearby travel routes. To address these challenges, each car is equipped with real-time traffic information for its immediate surroundings, allowing route adjustments at each time step.

The authors propose a mathematical model to represent the effects of earthquakes and traffic flows near exit nodes. The earthquake's impact is initially static, with edge weights (travel times) increasing at the beginning of the simulation. Subsequently, the ongoing dynamic effect of the earthquake and the traffic flow near exit nodes dynamically update edge weights as the car moves between nodes.

To assess the effectiveness of the strategy, a dataset with different problem instances is generated using the Python OSMnx package to convert real-world city maps into graphs. The study demonstrates the potential for efficient emergency evacuation route planning during natural disasters, considering both the earthquake's impact and evolving traffic conditions near exit points.

What does the research involve?

The proposed solution to the emergency escape routing problem involves a hybrid quantum neural network (HQNN) using SL. The HQNN iteratively selects the next node based on the car's current state and map. The SL model is trained on a dataset generated by node-wise Dijkstra's algorithm, approximating Dijkstra's algorithm while accessing only a limited map portion.

Input data for the SL approach includes earthquake coordinates, start and destination node coordinates, adjacent edges with their respective edge weights representing travel time, and edge betweenness centralities. Two heuristic indicators representing global information are introduced. Feature-wise linear modulation (FiLM) neural networks create a smooth and trainable conditional network, mitigating resource intensity and enhancing accuracy. The hybrid network (PHN) combines a classical NN and a variational quantum circuit (VQC), generating the likelihood of choosing neighboring nodes. The output of the HQNN predicts the next node for the evacuation route.

Findings of the study

The hybrid quantum approach to the shortest path problem was evaluated for improvement over classical machine learning. Results showed that the HQNN outperformed the classical neural network (NN), achieving an average accuracy of 94% compared to 87% for the NN. The HQNN also had a higher arrival rate of 95% compared to 87% for the NN. Furthermore, the HQNN made more robust choices in dynamically changing environments and found faster or equal paths to Dijkstra's algorithm. Practical analysis demonstrated the feasibility of running the model on ion-based quantum computers for short paths with few predictions. The ZX-calculus revealed reducible, trainable parameters, while Fourier analysis showed the model's expressivity. Finally, the Fisher information matrix provided insights into model trainability.

Conclusions

In summary, the present study explores supervised hybrid quantum machine learning for optimizing emergency evacuation routes during natural disasters. The hybrid approach outperformed the classical node-wise Dijkstra's algorithm by 7% in average accuracy. The quantum part of the hybrid model significantly contributed to the results. Future research should test the approach on different and larger graphs and explore other quantum machine learning techniques for similar optimization problems in dynamic environments, such as reinforcement learning. This work showcases the entire stack of hybrid quantum machine learning and its potential for real-world applications.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.