Decision Theory Research and Development

In this Phase I effort, Information Extraction and Transport, Inc. (IET) proposes to develop new representation formalism and inference algorithms for Partially Dynamic Decision Networks (PDDNs), which are a decision-theoretic extension of PartiallyDynamic Bayesian Networks (PDBNs) that have been developed by IET for the Ballistic Missile Defense Decision Architecture (BMD DA) and are extensively used by BMD DA developers. A PDBN extends a BN by introducing dynamic nodes whose values change overtime so that the models can include both static and dynamic nodes. PDBN solution algorithms exploit this additional structure in order to effectively compute the values of the joint probability distribution of the random variables in the PDBN based ondynamically observed evidence, such as sensor reports. By adding decision making capability to PDBNs, PDDNs will enable large-scale modeling of missile defense decision problems that involve dynamics (e.g., dynamically changing re-entry vehicle locations)and uncertainties (e.g., sensor noise) to build missile defense systems that use the developed models to make optimal decisions that maximize the expected current and future benefit by utilizing all information available through real-time streaming inputsfrom multiple sensors. Information Extraction and Transport, Inc. (IET) anticipates that the benefits of this Phase I effort will include: 1) a formalization of Partially Dynamic Decision Networks (PDDNs), a novel framework for dynamic decision makingunder uncertainty, and theoretical results regarding the correctness and computational complexity of inference algorithms, 2) a prototype implementation of a PDDN solution algorithm, and 3) a prototype PDDN compiler that compiles a PDDN specification togenerate procedural code that can be embedded in a real-time environment. IET anticipates that the Phase II implementation of these algorithms will support large-scale BMD PDDN modeling (for sensor management, RV/object discrimination and tracking, forexample) and real-time decision making with streaming sensor inputs. Beyond the BMD applications, our framework is applicable to a wide range of complex control problems. The potential applications include military, medical, and financial decision making,which all involve a complex model of the domain, making decisions in real-time in response to streaming inputs.