Improving COBRA Performance Using Context-adaptive Algorithms and Algorithm Optimization

This proposal addresses issues concerning operation of the Coastal Battlefield Reconnaissance and Analysis (COBRA) system in unseen locales and against unseen target types. The COBRA detection algorithms are based on statistical classifiers that are trained using samples from past data collects. As a result, the algorithms may not perform well against unseen situations. Further, once data have been collected in new situations the algorithms may need to be re-optimized, but this is currently a time-consuming effort that requires subject matter expertise. To mitigate these issues Aret? proposes a two-fold approach that aims to improve COBRA detection performance both in previously seen and in new, unseen locales. One aspect of this approach involves developing an offline, stand-alone application called the COBRA Optimization Tool (COBRA OT), which automatically determines the best values of algorithm parameters to provide optimal performance of the algorithms. The second thrust involves creating context-adaptive algorithms by identifying and exploiting information concerning the background settings (contexts) of data collects to improve local detection performance. The two aspects of Aret??s approach are complementary: context-adaptive algorithms will improve performance in each locale, and COBRA OT will optimize algorithm performance globally across all locales.