Multi-Agent Inverted Transformer for Flight Trajectory Prediction

TL;DR

The paper tackles multi-agent flight trajectory prediction by explicitly modeling intra-agent dynamics and inter-agent interactions. It introduces MAIFormer, a two-stage attention architecture with masked multivariate attention for each aircraft's trajectory and agent attention for social context, built atop an inverted-embedding tokenization scheme. On real ADS-B data from ICN arrivals, MAIFormer achieves state-of-the-art predictive accuracy and yields interpretable agent-level attention patterns, evidenced by lower attention entropy compared with prior multi-agent models. The approach holds practical promise for air traffic control decision support and invites future work on human-in-the-loop validation and integration of exogenous information such as weather and procedures.

Abstract

Flight trajectory prediction for multiple aircraft is essential and provides critical insights into how aircraft navigate within current air traffic flows. However, predicting multi-agent flight trajectories is inherently challenging. One of the major difficulties is modeling both the individual aircraft behaviors over time and the complex interactions between flights. Generating explainable prediction outcomes is also a challenge. Therefore, we propose a Multi-Agent Inverted Transformer, MAIFormer, as a novel neural architecture that predicts multi-agent flight trajectories. The proposed framework features two key attention modules: (i) masked multivariate attention, which captures spatio-temporal patterns of individual aircraft, and (ii) agent attention, which models the social patterns among multiple agents in complex air traffic scenes. We evaluated MAIFormer using a real-world automatic dependent surveillance-broadcast flight trajectory dataset from the terminal airspace of Incheon International Airport in South Korea. The experimental results show that MAIFormer achieves the best performance across multiple metrics and outperforms other methods. In addition, MAIFormer produces prediction outcomes that are interpretable from a human perspective, which improves both the transparency of the model and its practical utility in air traffic control.

Figures (8)

• Figure 1: Conventional approaches model inter-agent interactions by applying attention mechanisms across trajectory points, whereas MAIFormer models these interactions through direct agent-to-agent attention.
• Figure 2: Overall architecture of MAIFormer. The past air traffic scene is first tokenized into variate tokens. The sequence of variate tokens from all agents is fed into stacked MAIFormer layers where spatio-temporal and social interaction patterns of the air traffic scene are captured with masked multivariate attention and agent attention, respectively.
• Figure 3: Air traffic scene tokenization. We apply an inverted embedding to each agent’s past trajectory and generate a sequence of variate tokens from all agents. In this example, both the number of agents $N$ and the number of variates $F$ are set to three, which results in three tokens per agent and a total of nine tokens after tokenization. For consistency, we maintain these settings (three agents and three variates) in the subsequent figures.
• Figure 4: Illustration of masked multivariate attention (MMA) (left) and agent attention (AA) (right). In the MMA, the mask matrix $\mathbf{M}$ restricts the attention scores only to the variates from the same agent, which prevents information sharing across agents. In contrast, the AA allows all agents to share information with one another.
• Figure 5: Illustration of air traffic scene construction using a sliding time window.