COMMITTEE CHAIR: Dr. Lin Li
COMMITTEE CO-CHAIR: Dr. Lijun Qian

TITLE: EFFECTIVE TRAFFIC ROUTING UNDER DISASTROUS EVENTS THROUGH MACHINE LEARNING

ABSTRACT: Disastrous events such as hurricanes, wildfires, and floods routinely disrupt transportation networks, preventing emergency response, evacuation, and the delivery of critical supplies. Conventional optimization-based routing methods offer strong theoretical guarantees but cannot scale and are often too slow to support time-sensitive decision-making during disaster, while prior machine learning approaches to disaster routing have largely prioritized prediction accuracy over inference speed and data efficiency. This thesis investigates whether lightweight machine learning models can deliver both competitive routing accuracy and superior inference speed when trained on limited data, addressing a practical gap between existing research and real-world deployment needs. The road network is modeled as a synthetic abstraction of the Houston metropolitan highway system, comprising 22 nodes and 38 bidirectional edges, and the routing problem is formulated as a Minimum-Cost Flow problem. Disaster impact is simulated by systematically perturbing edge costs across twelve road groups, yielding 531,441 distinct disaster scenarios used to generate training and testing data. Lightweight machine learning models such as Random Forest with Regressor Chain, Bidirectional Long Short-Term Memory (BiLSTM), Transformer, and a policy-gradient Reinforcement Learning model, are benchmarked against classical optimization and approximation algorithms such as Linear Programming Relaxation, Successive Shortest Path, and Cost-Scaling Push Relabel. Results show that the BiLSTM model achieves the highest routing accuracy among the learning-based methods, reaching 75.26% correct routing with only 30\% of available training data, while the Reinforcement Learning model achieves the fastest inference time at a moderate accuracy trade-off. All machine learning models substantially outperform classical algorithms in inference speed, though the latter retain guaranteed optimality. These findings demonstrate that selective lightweight machine learning models can serve as practical, time-efficient alternatives for disaster-impacted traffic routing where rapid decision making is essential.

Keywords: Disaster Response, Minimum-Cost Flow, Traffic Routing, Machine Learning, Random Forest, Regressor Chain, Bidirectional Long Short-Term Memory (BiLSTM), Transformer, Reinforcement Learning, Data Efficiency, Real-Time Decision Making.

Room Location: S.R. Collins Building, Room 111L