Research Of Emergency Task Dynamic Planning Based On Reinforcement Learning

Public safety related to the fields of major national infrastructure and social security,which is the cornerstone of National security. However, China is suffered a huge lossbecause of the public emergencies, which seriously influences the country’s social stabilityand economic development. The emergency management of public safety has become oneof the important national policies. The fundamental task of emergency management is tomake quick and effective response to emergencies. However, the characteristics, complex,dynamic, uncertainty and strong time constraints, make the decision-making process inemergency management complicated. Emergency rescue process has become a complexdynamic task planning.A multi-task dynamic planning model of the emergency rescue process is built in thispaper according to the four characteristics. In this paper, not only the resource constraintsand time windows, but also the uncertainty, dynamic and complexity in the emergencyrescue process are considered to build an emergency rescue-MDP model. In this model, allpossible states, rescue operations and rescue strategy are considered to reflect thecharacteristics of the actual rescue in the maximum extent. The hierarchical reinforcementlearning algorithm based on Option is used to solve the model. This algorithm caneffectively solve the problem caused by uncertainty state and the huge state space. Withthe automatic generation and execution process of Option, the multi-task model inemergency rescue process is dynamically planned. Considering the alarm mechanism inemergency rescue process, the Interruption Option is designed which enhanced the rescuequality. In order to prove the validity of the model and algorithm, using a flood relief caseas the background, the emergency rescue algorithm based on Option is used in thesimulation. And to make a confrontation, the algorithm based on the traditional Q-learningis also used in the same model. By a large number of simulation experiments, thealgorithm based on Option is proved to be more significantly effective than the algorithmbased on traditional Q-Learning whether in the rescue time or rescue quality.