Evacuation route recommendation using auto-encoder and Markov decision process

Evacuation route recommendation plays an important role in emergency safety management, especially for natural disaster. When refugees flee a disaster area, the most important thing is to QUICKLY find the GLOBAL OPTIMAL evacuation route through analyzing the current situation in real time. Because the data for evacuation route recommendation is high-dimensional and huge-size, it is challenging to find an approach to quickly analyze such complex data collected from the current situation to find the optimal evacuation route. Most existing methods addressed this problem through analyzing a small part of the data (i.e., neighborhood) or reduced-size data, so that the important features of the data may not be retained. Therefore, this paper proposed a machine learning based method for evacuation route recommendation, which employs the auto-encoder method to reduce the data, and then conducts a reinforcement learning based route selection algorithm on the reduced data. Firstly, the feature-retained data reduction method is achieved through using the auto-encoder algorithm based on multilayer perception. By doing so, the complex high dimensional big data can be visualized in a 2D scatter plot, which can fully retain all the important features. This data reduction process is executed very efficiently, because an incremental training model is proposed. This model can also resolve the over-fitting problems caused by training the whole dataset together. Then, a Markov decision process based prediction model is proposed to design the global optimal evacuation route. Furthermore, new action rules, reward function, and discount factor have also been designed. Finally, the effectiveness of the proposed method has been demonstrated through analyzing evacuation routes using the meteorological data of Japan.