Hybrid Inferencing for Data Fusion and Situation Assessment

We propose to develop a hybrid inference system for data fusion and situation assessment for multi-source intelligence analysis with an application to Marine Corps operations in urban environments. Our proposed approach views an urban environment as a physical system whose state vector at any particular time is composed of a large number of both discrete and continuous variables representing properties of tracked entities. Inferencing on such vector-based models exploits both causal dependencies among variables in the state vector via its Dynamic Belief Network (DBN) representation and vector decomposition into weakly interacting subcomponents; such decomposition naturally occurs due to multiple tracked adversarial units working autonomously with intermittent communications amongst themselves. To efficiently utilize the decomposition, instead of straightforward particle filtering based inferencing, the proposed algorithm maintains factored particles over clusters of state variables. Because the clusters have far fewer variables than the entire state space, the variance resulting from maintaining cluster distributions is much smaller than that from maintaining the belief state as a whole. Based on the DBN structure of each cluster of variables, the algorithm samples discrete modes and approximates the continuous variables by a multi-normal distribution updated at each time step by the unscented Kalman filter.