Currently there are many DOD applications where warfighters are asked to make critical decisions based on environmental conditions that are highly complex and where there is incomplete knowledge of the local conditions. An example of such a situation is that of the theater commander who must deploy his C2/ISR assets such as communications and sensing platforms without complete knowledge of the local electromagnetic environment and its effect on his ability to maintain good information exchange and reconnaissance data for his forces. This type of situation falls into a broad class of problems where decision theory and complex physical models must interact for optimal performance such as investment analysis, weather prediction, and organizational dynamics. Such problems have been cast in the mathematical framework of "partial observability" where only some components of the environment are known. We thus we need to model the uncertainty of the environment and weigh our actions accordingly. The approach conventionally used for such optimization is a Partially Observable Markov Decision Process (POMPD) where we can model both our situational knowns and unknowns and come up with the best actions to take based on our model of what we know and do not know. We propose to develop a distributed computational framework that manages the complexity of such a process for large system optimization and provide an approach to parallelize and maintain operation for the system as more information and updates to our underlying environmental models change.
原文链接
分享
引用
