Investigation of the Effects of Correlation and Autocorrelation in Classifier Fusion with Non-Declarations

摘要

Air Force doctrine requires reliable and accurate information when striking targets. Further, this doctrine states that fusion should be utilized whenever possible to ensure the best possible information is conveyed; there is no specific guidance as to how to fuse this information. This thesis extends the research found in Leap, Bauer, and Oxley (2004) to include a non-declared class. The Identification system operating characteristic (ISOC) was adapted to allow for non-declarations both at the individual sensor level as well as the fused output level. A probabilistic neural network (PNN) was also used as a fusion technique. A cost function was developed that incorporated misclassification error as well as non-declaration rules. In addition, a heuristic was developed to find optimal rules through a likelihood ratio method. Finally, a sensitivity analysis was performed.