Automating NAVWAR Performance Analysis through Machine Learning
Kevin A Schaal, Mr. Paul C Manz, Dr. Thomas J Blenk Jr., Ekta R Patel, JPEO Armaments and Ammunition, Mrugesh Patel, Frank Fabian, Ian McMichael, Dr. Christian Minor, Rachel Cooper, MITRE
Location: Ballroom D
Date/Time: Thursday, Jun. 15, 9:35 a.m.
Positioning, Navigation, and Timing (PNT) systems and sensors are expected to operate in battlefield environments that are increasingly contested by adversaries, which are referred to as Navigation Warfare (NAVWAR) environments. DoD Instruction 4650.08 “Positioning, Navigation, and Timing and Navigation Warfare” defines the NAVWAR environment as “the expected physical, electromagnetic, and cyber conditions in which a Positioning, Navigation, and Timing (PNT) system operates.” As operating environments and PNT solutions become more complex in both commercial and military applications, there is a growing need for Command and Control (C2) systems to have accurate estimates of the expected performance of a given navigation sensor or system available in near-real time embedded in their workflow(s).
Current high-fidelity analysis methods are complex, often involving the integration of multiple detailed sensor models, navigation models, and three-dimensional environmental models. Due to their computational and analysis process complexity, these solutions are not well suited for automating an operational performance assessment of a given system in a NAVWAR environment within a timeframe that supports the operational pace of mission planning and execution. During mission planning, it is typically necessary to develop multiple courses of action that support the objective. Additionally, as the operation moves from the mission planning phase to the mission execution phase, the rapid changes in the NAVWAR environment that occur on the battlefield will require these analyses to be updated. While conducting NAVWAR analysis for a single course of action using the existing, detailed navigation models in current C2 systems might be reasonable, running multiple analyses in parallel in real-time requires a different approach.
This paper proposes the use of machine learning to reduce the computational and analysis process complexity of modeling a given PNT system in an arbitrary NAVWAR environment. Multiple embedding schemes are examined to translate an infinite three-dimensional data array describing the environment, PNT system, and mission parameters into a fixed-length input vector. Multiple machine learning variants are trained using the results obtained from the more complex, high-fidelity navigation models. Each iteration of machine learning model training considers different parameters in the embedding scheme to characterize performance as a function of the available NAVWAR environment inputs. This relationship informs both the minimal and the optimal awareness of the NAVWAR environment needed for an accurate performance assessment while also minimizing computational complexity.
The performance of each machine learning model variant is evaluated using a confusion matrix. This allows for the comparison between variants to account for not only their overall accuracy, but also for their respective rates of false positives and false negatives. In some applications, a false positive can be more acceptable to the Warfighter than a false negative. A false positive would still result in execution of the mission, which is equivalent to the current state in which no NAVWAR assessment is performed, while a false negative would preclude execution of what would been a successful mission (at least from a PNT perspective). Therefore, comparing each of the four quadrants of the confusion matrix for the model variants provides for a more comprehensive picture of their performance.
The proposed machine learning approach enables rapid PNT mission planning and accelerated decision making without increasing the cognitive burden on operators, requiring significant additional resources on C2 computing systems, or impacting existing operational timelines.
DISTRIBUTION STATEMENT A: Approved for public release; distribution unlimited. PAO Log #202-23, 03FEB23.