Abstract:In order to solving the problem of low intelligence of arrival flights scheduling, arrival flights scheduling reinforcement learning model was proposed. First of all, the states, actions, agent, environment, rewards, constraint and Q learning of the model were defined. The state of arrival flights scheduling reinforcement learning was the arrival time of each arrival flight. The action of the model was the adjustment of the arrival time. The arrival time was adjusted by agent, the environment responded to the action, then a new arrival time and reward value were passed to the agent. The delay times, delay cost and impact on subsequent flights were chosen as the rewards. At the same time, the model considered the flight can not advance landing, the distributed arrival time was not earlier than the schedule arrival time and the arrival flow of the airport can not exceed the arrival capacity of the airport. The model was validated using of Shuangliu airport real flight data. First come first service model and reinforcement learning model were compared and analyzed from sequence, delay time, delay cost, delay cost of subsequent flights and rewards. The rewards value of first come first service model was 3164. The rewards value of reinforcement learning model was 2880. Reinforcement learning model was superior to first come first service model. The evaluation index, weight, constraint can be adjusted according to air traffic control actual working conditions. The model provided decision-making for air traffic controller.