Sequential Modeling of Complex Marine Navigation
This AAAI paper applies sequential learning to predict and model vessel behavior in complex waterways using historical AIS data, enabling anomaly detection and route optimization at operational scale. The work demonstrates that reliability methods developed in robotics transfer directly to maritime operations - a domain where AI errors carry significant safety and regulatory consequences. Published at AAAI 2024.
Marine navigation in confined waterways - ports, channels, fjords - involves complex multi-agent dynamics where vessel trajectories must account for traffic separation schemes, weather conditions, tidal effects, and the intentions of other vessels. Automating analysis of these dynamics requires models that capture long-range temporal dependencies in trajectory data.
Data and Setting
We work with Automatic Identification System (AIS) data: GPS position reports broadcast by commercial vessels at regular intervals. AIS data is rich in trajectory information but noisy - position errors, signal dropouts, and GPS drift are common. The modeling pipeline must be robust to these data quality issues.
The target domain is a section of complex Canadian coastal waterway with high ferry and cargo traffic density, providing a realistic operational challenge.
Sequential Model
The core model is a transformer-based sequence encoder that takes a window of recent AIS observations and predicts the distribution over future positions. Key design choices:
Uncertainty quantification: The model outputs a distribution over future positions rather than a point estimate. This enables the downstream anomaly detector to flag vessels whose observed trajectory is improbable under the model - distinguishing genuine anomalies from model uncertainty.
Multi-vessel context: The encoder conditions on nearby vessel trajectories, enabling the model to capture interaction dynamics (give-way situations, overtaking, anchoring).
Anomaly detection: We define anomalous navigation as trajectories that receive consistently low likelihood under the model. Validated against a manually labeled set of near-miss incidents from maritime authority logs.
Results
The model achieves strong predictive accuracy on held-out trajectory sequences and successfully identifies a high proportion of the manually labeled incident cases as anomalous, with a low false positive rate that would be operationally acceptable for human-reviewed monitoring.