Large-scale Passenger Behavior Learning and Prediction in Airport Terminals based on Multi-Agent Reinforcement Learning

Abstract

For the problem of predicting passenger flow in airport terminals, multi-agent reinforcement learning is applied to airport terminals simulation. Multi-Agent Reinforcement Learning based on Group Shared Policy with Mean-field and Intrinsic Rewards (GQ-MFI) is proposed to predict passenger behavior in order to simulate the distribution of flow in different areas of the terminal at different time periods. Independent learning of multi-agent may lead to environmental instability and long convergence time. To improve the adaptability of agents in non-stationary environments and accelerate learning time, a multi-agent grouping learning strategy is proposed. Clustering is used to group multi-agent, and a shared Q-table is set within each group to improve the learning efficiency of multi-agent. Meanwhile, in order to simplify the interaction information among the agent after grouping, the idea of average field is used to transmit partial global information among the agent within the group. Intrinsic rewards are added to make the agent closer to human cognition and behavioral patterns. By conducting the airport terminal simulations using Anylogic, the experimental results show that the training speed of this algorithm is 17% higher than that of Q-learning algorithm, and it achieves good prediction accuracy in predicting the number of security check passengers with a time scale of 10 minutes.