A computer program estimates great-circle-arc (GARC) flightpaths of multiple aircraft, given time-tagged point locations from radar sightings of the aircraft. The problem of generating such estimates is called a "clustering" problem, and is solved in this instance by use of a neural-network clustering algorithm. Points are considered in pairs; the likelihood that any two points are on the flightpath of the same aircraft is quantified by an "association value" based on the flight dynamics of the aircraft. The program implements a Boltzmann machine, the sparse architecture of which provides for only partial satisfaction of the constraints of a cost function; this, together with a special graphical interface, serves as an aid in determining GARCs. The neural-network algorithm operates on all points simultaneously and performs a global optimization through simulated annealing; thus, it is in many instances superior to both traditional clustering algorithms that operate on points sequentially, and to other neural-network algorithms that perform local optimization. The neural network can also readily be implemented in hardware.
This work was done by John Spagnuolo, Jr., of Caltech for NASA's Jet Propulsion Laboratory. NPO-20288
This Brief includes a Technical Support Package (TSP).
Program computes flight paths from point radar sightings
(reference NPO20288) is currently available for download from the TSP library.
Don't have an account?
Overview
The document discusses a program developed by John N. Spagnuolo, Jr. at NASA's Jet Propulsion Laboratory, which computes flightpaths from point radar sightings. This innovative approach utilizes a neural network clustering algorithm based on a Boltzmann machine architecture to effectively determine Great Arc Circles (GARCS), which represent the paths of aircraft on the surface of an imaginary sphere surrounding the Earth.
The primary challenge addressed in the document is the clustering problem, where multiple radar sightings generated by various sensors need to be grouped into sets corresponding to individual aircraft. The goal is to connect these sightings in a time-ordered manner to form accurate flightpaths. The proposed solution employs a sparsely connected neural network, referred to as a neural clusterer, which is designed to converge quickly to optimal or near-optimal solutions.
The document highlights the unique aspects of the neural clusterer, including its integration with a non-neural tracker, which simplifies the network's architecture and allows for future experimentation across multiple processors. The Boltzmann machine architecture facilitates the search for optimal solutions, making it suitable for various applications beyond aviation, such as in space and oceanic domains.
Results from simulations demonstrate the effectiveness of the neural network in accurately identifying GARCS, achieving a 98% match between the tracker output and real-world data. Although the document does not rigorously prove that the solutions reached are globally optimal, it indicates that the architecture consistently produces high-quality results, especially in scenarios where traditional non-neural methods can easily determine optimal solutions.
The document also discusses the potential for further research, including the use of supercomputers and neurocomputers to enhance the clustering process. It suggests that multiple copies of the neural clusterer could operate in parallel, each working on different values of the clustering parameter, to improve convergence properties and accuracy.
In summary, this document presents a significant advancement in flightpath computation through the application of neural networks, showcasing the potential for improved accuracy and efficiency in tracking aircraft using radar sightings. The findings encourage further exploration of neural network applications in various fields, emphasizing the versatility and effectiveness of this technology.
