Multi-agent learning for data-driven air traffic management applications

Research in developing data-driven models for Air Traffic Management (ATM) has gained tremendous interest in recent years. However, data-driven models are known to have long training time and require large datasets to achieve good performance, and the majority of proposed data-driven models ignores ATM system's multi-agent characteristic. To fill the research gaps, this paper proposes a Multi-Agent Bidirectional Encoder Representations from Transformers (MA-BERT) model, which fully considers the multi-agent characteristic of the ATM system and outputs results based on all agents in the airspace. Additionally, compared to most data-driven models that are designed for a single application, the proposed MA-BERT's encoder architecture enables it to be pre-trained through a self-supervised method and fine-tuned for a variety of data-driven ATM applications, saving a substantial amount of training time and data usage. The proposed MA-BERT is tested and compared with other widely used models using the Automatic Dependent Surveillance-Broadcast (ADS-B) data recorded in three airports in South Korea in 2019. The results show that MA-BERT can achieve much better performance than the comparison models, and by pre-training MA-BERT on a large dataset from a major airport and then fine-tuning it to other airports and applications, a significant amount of the training time can be saved. For newly adopted procedures and constructed airports where no historical data is available, the results show that the pre-trained MA-BERT can achieve high performance by updating regularly with small amount of data.