Clustering by means of a Boltzmann machine with partial constraint satisfaction

Abstract

The clustering problem refers to the partitioning of target sightings into sets. Two sightings are in the same set if and only if they are generated by sensor detections of the same target and are in the same great circle arc (GARC) trajectory of that target. A Boltzmann machine is developed whose sparse architecture provides for only partial constraint satisfaction of the associated cost function. This together with a special graphics interface serve as an aid in determining GARCs. Our approach differs from others in that the neural net is built to operate in conjunction with a non-neural tracker. This further restricts the architectural complexity of the network and facilitates future experimentation regarding decomposition of the neural net across several Von Neumann processors. Also, the Boltzmann machine architecture eases the effort of finding optimal or near optimal solutions. Results are presented. The demonstrated feasibility of neural GARC determination encourages investigation into the extension of its role in the track formation process utilizing an environment that includes supercomputers, neurocomputers, or optical hardware. The network architecture is capable of identifying a host of geometric forms other than GARCs and can thus be used in several domains including space, land, and ocean.