Causality-Aware Machine Learning for Path Correction

Abstract

Path identification and prediction are key challenges in many modern aerospace applications. Path identification has broad applications to both the routing and tracking of systems and is crucial to the navigation of autonomous systems. Machine learning (ML) has proven to be effective at learning complex patterns and relationships in tracking data and using them to make sophisticated predictions about future entity motion. However, ML agents performing these tasks typically rely on knowledge of the past motion of entities to help predict their future locations. While the true path traversed by an entity in motion may be used in simulation studies, in practice this is not the case. Position and motion data of entities is often estimated from information collected by sensors that have inherent error sources. These error sources impede a system’s ability to accurately track entity motion. As motion tracking and path prediction are recursive processes, the errors can be further amplified and compounded over time. Existing approaches to correct for track errors generally focus on either solely correcting current measurements or use overly simplistic interpolation to correct for past errors, both of which often fail to maintain sufficient track integrity to allow for accurately predicting future states. Here, we propose a new method for repairing tracks based on causality-aware ML through the evaluation of counterfactual scenarios. Approaching track repair in this way allows a system to examine suspect sections of a track in their entirety and correct errors in them in a way that is consistent with both an entity’s location and its past motion, producing significantly improved results. We have implemented this approach to create a track repair tool using a combination of recurrent neural networks (RNNs) to predict target motion and the non-dominated sorting genetic algorithm II (NSGA-II) to identify and evaluate counterfactual paths. We have evaluated the performance of our track repair tool on simulated motion data and found that it drastically reduces track errors and improves predictions of an entity’s future location for multiple time steps.