An Efficient Computational Approach to Target Discrimination
Department of Defense
Missile Defense Agency
Agency Tracking Number:
Solicitation Topic Code:
Small Business Information
6 New England Executive Park, Burlington, MA, 01803
Socially and Economically Disadvantaged:
Vice President, SD Divisi
Vice President, SD Divisi
AbstractTarget discrimination is one of the critical components of the midcourse phase processing in providing effective missile defense. Currently, detailed models that use an elaborate discrete Bayesian network predict the sensor phenomenology for each possibletarget type and sensor. The predicted sensor phenomenology is compared to the observed phenomenology, and the target type estimate is updated. The complex relationships modeled in the Bayesian network is extremely computationally expensive and cannotkeep up with real time. We propose to use a Gaussian sum approximation technique to model the continuous variate probability density functions (PDFs) that occur in the Bayesian network used in target discrimination. Gaussian sum approximations have thepotential of increasing the accuracy of estimating the continuous valued PDFs that are currently modeled by picewise linear discrete functions. We will address both the accuracy of the Bayesian network and, perhaps more importantly, the computationalefficiency of the Gaussian sum approach in updating the node PDF when new sensor information is received-only eight computations are needed to update a term in the Gaussian sum. In summary, we will develop and evaluate the computational efficiency ofGaussian sums to model continuous components used in the Bayesian network for target discrimination modeling. The significance of developing a real-time updating of the Bayesian network used for target discrimination is increased performance in themidcourse phase. This increase in the performance of discriminating between warheads and decoys significantly increases the likelihood that warhead threats will be killed by space-based interceptors, as well as increasing the accuracy of thetarget-clutter map used in the handoff to ground-based interceptors in the terminal phase. Potential commercial applications of computationally efficient Bayesian network models include modeling complex financial networks that provide buy and sell signalsfor stocks, in addition to supplying real-time models used to detect and select automated responses to sophisticated attacks on computer network systems.
* information listed above is at the time of submission.