Design and construction of normative risk management and decision systems.

摘要

My thesis research extends existing approaches and develops new techniques for the design and construction of normative decision software systems. Such systems, which are based on the normative principles of decision analysis, are used to formulate and evaluate diagnostic and decision models, typically as influence diagrams. Problem formulation and evaluation techniques are applied to engineering risk management situations, which are a class of partially observable sequential decisions.; Engineering risk management situations exhibit a characteristic dynamic pattern. Following an initiating event, a physical system degrades. Warning signal thresholds balance false alerts and lead time for response to an abnormal situation. Delay between signal and action balances system degradation and information gathering. Alternative actions balance cost and improvement.; I decompose the engineering risk management problem into four basic activities--diagnosis, dynamic evolution, decision making, and information gathering--and corresponding representations--belief network, Markov model, influence diagram, and decision tree. I devise an automatic problem formulation method for sequential information gathering that performs flexibly under resource constraints. The method explicitly considers temporal and other information-gathering alternative eligibility conditions to incrementally formulate an asymmetric set of test sequences. The method also handles the dynamics of learning (e.g., diagnostic tests) and of the underlying physical system state. I illustrate these techniques with rapid detection of and response to gas leaks on offshore oil platforms.; I design a domain-independent intelligent decision system that balances guidance and flexibility in assisting a user with constructing an influence diagram representation of a decision problem. I also design a method for customizing this domain-independent interactive system for a particular domain. The customization method enables decision analysts with domain expertise to build prototype intelligent decision systems whereas only designers skilled in both decision analysis and software programming can do so now. I illustrate these concepts with interactive formulation and evaluation of repair timing decisions following structural damage to an offshore platform.