Graphical models for planning, information integration and decision making

Information analysis and decision making are complex processes given that the sources of information are diverse and distributed, the acquired information is noisy, dynamic, incomplete and uncertain. The key issues include: identifying and sharing valuable information in a timely and efficient manner; integrating large volumes of disparate information to support better decisions; suggesting effective and robust courses of action (COA) for the targeted mission based on partial information; optimizing a team decision in a noisy and distributed environment. ^ In this thesis, we utilize multiple graphical models, including static Bayesian networks (BNs), dynamic Bayesian networks (DBNs), hidden Markov models (HMMs) and influence diagrams (IDs), to address the problems of stochastic planning, information integration (fusion), and distributed decision making. A new methodology, which combines DBN and genetic algorithms, is introduced to obtain a near-optimal robust strategy for planning in a stochastic environment. The problem is formulated as a DBN-based stochastic mission model, while a genetic algorithm is utilized in an outer loop to search for a near-optimal strategy with DBN inference serving as a fitness evaluator. A collaboration scheme which constructs BNs and HMMs in a hierarchical fashion is discussed for fusing uncertain and diverse information from multiple agencies. HMMs function in the bottom observation layer to process the dynamic transactions and detect suspicious activities, if any. BNs act in the top layer to integrate local information and obtain a global assessment for the scenario being monitored. Influence diagrams are widely accepted for problems, where a single decision maker is making sequential decisions under uncertainty. However, real world situations are very likely to involve parallel decisions, where multiple decision makers cooperate to arrive at a team decision. Iterated influence diagrams, which extend traditional influence diagrams, employing a person-by-person optimization scheme, are developed and are used to solve a parallel distributed hypothesis testing problem.