Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multi-Agent Based Transfer Learning for Data-Driven Air Traffic Applications

Chuhao Deng, Hong-Cheol Choi, Hyunsang Park, Inseok Hwang

TL;DR

MA-BERT introduces agent-aware attention to explicitly capture multi-agent interactions in air traffic management and pairs it with a pre-training and fine-tuning transfer-learning framework. Trained on a large ADS-B dataset from RKSI and transferred to RKSS and RKPK, the approach achieves superior trajectory and ETA predictions while dramatically reducing total training time compared to training from scratch. The results show strong cross-airport generalization, with incremental learning enabling adaptation to new airports with little or no historical data. This framework offers a scalable, data-efficient path to deploying data-driven ATM applications across diverse airports and procedures.

Abstract

Research in developing data-driven models for Air Traffic Management (ATM) has gained a tremendous interest in recent years. However, data-driven models are known to have long training time and require large datasets to achieve good performance. To address the two issues, this paper proposes a Multi-Agent Bidirectional Encoder Representations from Transformers (MA-BERT) model that fully considers the multi-agent characteristic of the ATM system and learns air traffic controllers' decisions, and a pre-training and fine-tuning transfer learning framework. By pre-training the MA-BERT on a large dataset from a major airport and then fine-tuning it to other airports and specific air traffic applications, a large amount of the total training time can be saved. In addition, for newly adopted procedures and constructed airports where no historical data is available, this paper shows that the pre-trained MA-BERT can achieve high performance by updating regularly with little data. The proposed transfer learning framework and MA-BERT are tested with the automatic dependent surveillance-broadcast data recorded in 3 airports in South Korea in 2019.

Multi-Agent Based Transfer Learning for Data-Driven Air Traffic Applications

TL;DR

MA-BERT introduces agent-aware attention to explicitly capture multi-agent interactions in air traffic management and pairs it with a pre-training and fine-tuning transfer-learning framework. Trained on a large ADS-B dataset from RKSI and transferred to RKSS and RKPK, the approach achieves superior trajectory and ETA predictions while dramatically reducing total training time compared to training from scratch. The results show strong cross-airport generalization, with incremental learning enabling adaptation to new airports with little or no historical data. This framework offers a scalable, data-efficient path to deploying data-driven ATM applications across diverse airports and procedures.

Abstract

Research in developing data-driven models for Air Traffic Management (ATM) has gained a tremendous interest in recent years. However, data-driven models are known to have long training time and require large datasets to achieve good performance. To address the two issues, this paper proposes a Multi-Agent Bidirectional Encoder Representations from Transformers (MA-BERT) model that fully considers the multi-agent characteristic of the ATM system and learns air traffic controllers' decisions, and a pre-training and fine-tuning transfer learning framework. By pre-training the MA-BERT on a large dataset from a major airport and then fine-tuning it to other airports and specific air traffic applications, a large amount of the total training time can be saved. In addition, for newly adopted procedures and constructed airports where no historical data is available, this paper shows that the pre-trained MA-BERT can achieve high performance by updating regularly with little data. The proposed transfer learning framework and MA-BERT are tested with the automatic dependent surveillance-broadcast data recorded in 3 airports in South Korea in 2019.
Paper Structure (13 sections, 15 equations, 8 figures, 7 tables)

This paper contains 13 sections, 15 equations, 8 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. MA-BERT
  4. Pre-Training Scheme
  5. Fine-Tuning Scheme
  6. Results and Analysis
  7. Data Description and Preparation
  8. Downstream Tasks and Evaluation Metrics
  9. Neural Network Setting
  10. Model Comparison
  11. Fine-Tuning Results
  12. Incremental Learning Results
  13. Conclusion

Figures (8)

  • Figure 1: General framework of pre-training and fine-tuning
  • Figure 2: Proposed MA-BERT structure
  • Figure 3: An illustration of the mask in agent-aware attention
  • Figure 4: An illustration of the agent-aware attention weight matrix
  • Figure 5: An example of a masked scene for pre-training the MA-BERT: dashed parts are the masked parts
  • ...and 3 more figures