Road reconstruction during post-earthquake crises is vital for search operations, relief efforts, the movement of individuals from affected areas, and the sending of food and relief items. Blockage of the roads may prevent the delivery of critical goods or cause a lack of cars and ambulances access to the affected areas to evacuate injured people. In many studies, road debris removal is a sophisticated issue, so only a few roads were considered or aimed at finding a route through which all affected areas can be visited. Since roads have limited capacity for vehicle traffic during severe earthquakes, a lack of full debris removal may lead to delayed movement of injured people from affected areas, and debris left after the crisis may result in environmental and psychological harm to affected people. To solve this problem, a two-step model-free approach based on reinforcement learning methods is developed for full debris removal from damaged roads. In the first step, the damaged area is initially assessed using an unmanned aerial vehicle (UAV) and a state-action-reward-state-action (SARSA) algorithm to estimate the damage. In the second step, debris removal groups remove debris from all blocked roads based on the information reported by the UAV. For this purpose, a heuristic algorithm is developed based on debris removal groups, the divergence of the genetic algorithm is ensured and the agent training matrix is regulated using the neural network. The proposed solution methods are examined under three scenarios of severe, moderate, and weak earthquakes in Rudbar city as an artificial intelligence application. The results indicate that the debris removal operation is completed in less than 80 h in the severe case if the proposed method is used, whereas the algorithm achieves the proper solution in less than 45 min.
