Adaptive Data Fusion Technology

The prodigious amount of information provided by surveillance systems and other information sources has created unprecedented opportunities for achieving situation awareness. However, due to the fact that Air Force missions are constantly evolving and specific user information needs are not predictable in advance, it is necessary to devise fusion control strategies that are adaptive to user needs and the context of each mission. However, the optimal control problem arising from the desired adaptive control capabilities is enormously complex. We propose to develop a methodology, called Neuro-Dynamic Programming, that combines elements of dynamic programming, simulation-based reinforcement learning, and statistical inference techniques to address the complex adaptive fusion control problem. The principal advantage of the Neuro-Dynamic Programming methodology is that the cost-to-go function that drives the optimal control strategy is not required to be computed analytically. Rather, the system learns the cost-to-go function via simulation-based reinforcement learning. In related work, Neuro-Dynamic Programming controllers have been shown to provide near-optimal control of complex systems with modest amounts of simulation-based training. Extending Neuro-Dynamic Programming for adaptive fusion control holds the promise of providing fusion capabilities that are responsive to user needs and sensitive to context dependencies of the performance of underlying fusion processes.