Abstract
A stalwart of navigation and simulation, Dead Reckoning (DR) is a method which calculates an entity's current position by extrapolating from its previous state using known speed, heading, and elapsed time. However, traditional DR approaches rely on the assumption that incoming data streams are accurate and trustworthy, often resulting in unrealistic behaviour, lack of interactivity, and catastrophic vulnerability to data poisoning. These limitations compromise immersion, reduce training effectiveness, and expose systems to potential manipulation.
We propose a novel approach which integrates Artificial Intelligence (AI) and Machine Learning (ML), particularly Reinforcement Learning (RL), to enhance the realism, autonomy, and resilience of simulated entities. By leveraging historical AIS data and procedurally generated maritime environments, we aim to train RL agents capable of navigating complex scenarios with greater situational awareness and responsiveness. The proposed system architecture includes a simulation environment built using the Gymnasium framework for the integaration of a convolution-based perception model, an anomaly detection layer to mitigate the effects of poisoned or unreliable data, and common interfaces to live AIS data, real-world terrain, and distributed simulation environment protocols for interoperability. We outline an experimental methodology using Deep Q-Networks (DQN), with the potential to scale to more advanced algorithms.
Evidence suggests that this combination of AI/ML with DR principles may yield improvements in realism, interactivity, and responsiveness, enhancing training and ensuring operational dominance in synthetic environments. Our proposed approach aims to produce digital shadows which maintain intelligent, context-aware behaviour even in the presence of data dropouts or latency, enabling entities to navigate plausibly and avoid unrealistic actions such as traversing terrain or disregarding battlespace constraints.