Knowledge representation and inference in intelligent decision systems
Jack Breese · 1987
Decision making in complex domains typically consists of a series of related decisions, each of which requires a different level of support and richness of representation. Rule based expert systems for decision support have been successful for well structured, well understood decision situations. As uncertainty increases and the preferred solution depends on the specific beliefs and preferences of an individual decision maker, more powerful techniques based on single person decision theory can be brought to bear. This research focuses on alternative means of representing and using knowledge regarding decision situations in a computer-based decision aid. A unified characterization of knowledge and inference for logical, probabilistic, and decision-theoretic reasoning is developed for intelligent decision support over a wide spectrum of decision situations. A representation of a decision domain consists of structures for representing the decision choices, alternative possible states or outcomes which might occur, the relationships between choices made and outcomes realized, and preferences of the decision maker for the various outcomes. The components are captured by an extension to first-order predicate logic in which propositions are used to represent states, alternatives, and objectives. Sentences in the language denote logical, probabilistic, or informational influences among propositions. A set of propositions and influences comprise a declarative description of a decision domain. Inference is the process of deriving new conclusions from the decision domain knowledge base. The inference techniques in this thesis focus on the process of constructing probabilistic and decision theoretic models from a declarative description of the domain. Deterministic, logical methods are integrated with probabilistic and decision-theoretic reasoning, allowing selection or combination of methods with the appropriate power and flexibility to solve a given problem. A prototype of the system, ALTERID, running on a high resolution LISP workstation, has been developed. The functionality of the system is demonstrated on an exemplary problem from the domain of financial securities trading.