How to Train Your Ship Traffic Model

Abstract

The proposed Houston Ship Channel Gate Complex (HSCGC) near Galveston, Texas, is a central element of the Texas coastal protection plan and is intended to reduce societal and economic risks from hurricanes and sea-level rise for the Greater Houston Port System. At the same time, the gate introduces a major new intervention in a heavily used waterway, with previous studies indicating that the proposed layout may create navigational hazards and act as a chokepoint that affects traffic well beyond the immediate gate location. This thesis situates the HSCGC as an example of a broader trend: increasing implementation of flood-protection and other fixed structures in navigable waterways that must continue to accommodate growing maritime transportation demand, raising the need for tools that can evaluate traffic impacts early in design.

The objective of this research is to contribute to the development of data-driven microscopic maritime traffic simulation models as a design tool for new maritime infrastructure, using the HSCGC as the case context. The main research question asks: "What requirements and characteristics must such a data-driven maritime traffic simulation model have to assess the impact of prospective maritime infrastructures on maritime traffic patterns in the context of designing the HSCGC?"

To answer this question, a mixed methodology is applied. A literature review establishes the state of the art in microscopic ship-traffic modelling and motivates the selection of AIS-based learning approaches, because they can reproduce complex manoeuvring behaviour without fully prescribing rules or equations. However, because purely data-driven models generalize poorly to unseen infrastructure, the thesis justifies the exploration of Safe Reinforcement Learning extensions, specifically safety filtering layers that can enforce collision and obstacle-avoidance constraints while deviating minimally from learned behaviour. Empirically, AIS data from 2024 is processed into trajectories to derive baseline traffic structure, kinematics, and interaction hotspots, while semi-structured interviews with expert navigators complement AIS by identifying operational constraints and anticipating behavioural changes under an HSCGC scenario. Finally, the selected simulator (ShipNaviSim) and extensions are evaluated on historical realism and situational adaptability using trajectory- and behaviour-focused performance indicators.

Results show that maritime traffic in the study area is highly structured yet interaction-rich: dominant channel-aligned flows coexist with frequent crossings (notably the Galveston-Point Bolivar ferry corridor), producing localized encounter hotspots and heterogeneous manoeuvring demand. The evaluated data-driven simulator reproduces goal-seeking motion and qualitatively plausible transit classes, but does not consistently match observed kinematic distributions (speed, drift, curvature, and acceleration), limiting quantitative realism. Among tested extensions, intermediate goals improve channel-following substantially, while an MPC-based safety layer reduces obstacle entry violations and supports scenario execution under modified geometries, though robustness remains challenging in head-on and high-density encounters.

The thesis concludes that a design-capable data-driven maritime traffic model must be a validated microscopic multi-agent AIS-driven simulator that reproduces site-specific route structure, interaction dynamics, and vessel heterogeneity, while explicitly accepting scenario inputs for new obstacles and changing demand patterns. Critically, it must incorporate a robust safety mechanism, such as safety filtering within a safe reinforcement learning framework, to enable credible and safe behaviour in previously unseen infrastructure configurations.

https://github.com/TUDelft-CITG/traffic-behaviour-cloning