A Deep Reinforcement Learning Evolution of Emergency State during Traffic Network

The organization and management of emergency scenarios is an important task that could prevent a huge loss of lives and property. Fast and reliable decisions would minimize the impact of the accident. Current decision making mainly relies on the experience of emergency personnel. It is hard to make suitable decisions immediately according to the situation. The appropriate decisions making always spend lots of valuable time. With the development of artificial intelligence, the learning-based algorithm, which enables an agent to learn from the environment, might compensate the subjective judgement and reduce the decision-making time. The reinforcement learning (RL) enables an agent to learn appropriate action from the environment. In this paper, a model is constructed using Deep Deterministic Policy Gradient (DDPG) to determine which measures would be taken to disposal accident quickly and efficiently when an accident occurs. The comparisons of the fire scene and non-fire scene are provided. This study also could be utilized in another emergency effect, and it is of great potential for coordinated decision-making throughout the emergency disposal.