Optimizing post-disaster road restoration with reinforcement learning: A traveler-behavior-aware approach

Abstract

This paper introduces an AI-driven method to optimize short-term road network restoration in disaster-affected areas, aiming to maximize post-disaster traffic acceleration. Our approach considers travelers' behavior, gradual adaptation to network changes, limited recovery resources, and uncertainties in recovery times. Addressing these complexities requires a stochastic approach for uncertainties, sequential decision-making for resource management, and a model-free technique for simulating traveler adaptation.

To tackle these challenges, we develop the Traveler-Adaptive Restoration Mechanism (TARM), integrating Reinforcement Learning (RL), the Markov Decision Process (MDP), and optimization-based day-to-day traffic simulation. The method is evaluated on Sioux Falls' road network under tornado scenarios based on historical data. Results highlight the influence of travelers’ route choices and the speed of restoration information dissemination on optimal policies.

Findings reveal that accelerating the road restoration process by increasing restoration resources does not necessarily enhance the traffic movement efficiency in disaster-affected communities during the disaster response period. Furthermore, we demonstrate that contrary to exiting studies, shortening restoration period is not an appropriate measure of efficiency for post-disaster restoration operations. In fact, reducing the restoration period may adversely impact traffic movements during the response phase in post-disaster situations.